Data refresh

What each refresh found, and what it did not publish.

ModelTree's dataset is refreshed by agents against primary sources, reviewed by an independent three-rubric panel, and gated deterministically. Every run is recorded here in full — including the runs that published nothing.

A run's working state is never committed. This page transcribes the durable record: the pull request body and the summary issue, both linked from each entry.

Runs recorded: 30
Pages fetched: 4,308
Claims proposed: 4,607
Edits published: 3,698
Items withheld: 605

Showing 1–10 of 30 runs

2026-09-24-83cd25

Re-verified 397 records across 43 creators; the panel refused four Anthropic and NVIDIA claims

Scope requested: All 44 creator records in web/src/data/organizations.json, each scouted against the primary channels catalogued for it in web/src/data/sources.json; 296 URLs fetched in a single sweep, 289 of which returned a body that was saved to disk and hashed over the exact bytes received. Changed web/src/data/sources.json, web/src/data/organizations.json, web/src/data/families.json and web/src/data/releases.json, and nothing else: 397 insertions against 397 deletions, every one of them a lastCheckedDate or verifiedAt value. No record was added, removed or renamed, so all four documents hold exactly the record counts they held at the anchor, and the four files are byte-for-byte the same length after the run as before it. web/asset-budgets.json was not re-recorded: the measured route:home figure did move, by one byte, which is discussed in the caveats, and one byte against the 2% drift allowance that file records fails nothing and changes no ceiling.PublishedSep 24, 2026397 edits posted · 46 items withheld

Swept all 44 creators against their own catalogued primary channels and found no new model, family or release anywhere: every one of the 441 claims this run proposed is a re-verification, moving a lastCheckedDate or verifiedAt to 2026-09-24 on a record that already existed. 397 of them were applied. The interesting result is the 44 that were not. Two rubrics independently rejected the Anthropic Sonnet 4.5 and Sonnet 4.6 announcement sources after noticing that both pages, fetched separately and returning different byte counts with no redirect, yielded the identical passage about a new life sciences research group -- a site-wide promotional component, not either article body, and so evidence that the fetched page is still the source the dataset names. The editorial rubric separately refused NVIDIA Cosmos as a family, because NVIDIA's own profile warns that Cosmos is a World Foundation Model Platform and that platform names are not family names. gate-evidence then refused the whole anthropic and nvidia bundles over those four claims, taking 40 unanimously accepted siblings down with them; that refusal was obeyed rather than cherry-picked around, which is why this entry records 441 claims proposed and 397 applied.

PreflightRanClean tree at merge-base 64f70ef9, gh authenticated, and no other open pull request touching a web/src/data path. npm was probed in both bare and .cmd form: the bare form is refused by the PowerShell execution policy while npm.cmd returns 11.9.0, which is installed-and-blocked rather than absent. The execution policy was not changed to make the bare form work, and reporting it as missing would have sent the next reader somewhere different. Dependencies were restored with npm ci, never a bare install, and git status --porcelain was empty afterwards, which is the check that the lockfile ADR 0004 protects was not rewritten.
ScoutRan296 fetch attempts covering every catalogued source; 289 returned a body and 7 returned HTTP 403, each recorded as a degraded channel against the creator it belongs to rather than folded into a flat error list. No browser user-agent was spoofed to get past any of them. Every body was written to disk and hashed over the bytes received, so every quote published in the pull request is checkable against the artefact it came from. A separate discovery pass over 543 candidate names found no uncatalogued model: all 543 resolved to historical checkpoints or size variants already represented. The scout output was then re-audited by an independent script that re-hashed every body and re-checked every quote for verbatim containment by a different normalisation route than the extractor used -- 441 claims and 672 evidence entries, zero failures -- and that audit was itself controlled by injecting three defects into a copy, all three of which it caught.
ReviewRanThe 441 claims were split into nine batches of at most 50 and put to three rubrics -- provenance, consistency, editorial -- as 27 independent blind reviews. No reviewer saw another rubric's verdicts, the scout's reasoning, or the chair's expectations, and each brief carried the ADR 0017 guardrail verbatim: lastCheckedDate and verifiedAt are claims that a person looked, and a link checker is not a person. 1,323 verdicts came back. The chair tallied them and did not vote: 437 claims met their own threshold and 4 did not. An earlier version of the tally script compared a numeric accept count against the threshold recorded as the string "2-of-3", which is always false, and duly reported all 441 claims rejected; that uniform answer was caught because a panel rejecting everything is not a finding, and the repaired script cross-checks the parsed number against the policy rather than trusting either alone.
GatesRangate-source-approval passed on all 45 bundles, reporting 273 inherited sources and zero proposed ones -- this run introduces no source, so there is no origin here that nobody has approved. gate-evidence passed 43 bundles and refused 2. gate-dataset, the full web validate suite and ci-preflight were run after applying. The one validate failure seen before committing was licence-identity.test.ts asserting that git status --porcelain src/data is empty, which it cannot distinguish from a tree that was already dirty for a legitimate reason; it passes once the dataset change is committed, which is the state CI actually evaluates.
PublishRanOne dataset commit and one ledger commit on a single branch, published as a pull request with auto-merge enabled so that web-ci gates the merge rather than any judgement of this run's. No direct push to main, no --admin, no --force. The dataset commit precedes the ledger commit deliberately: refreshLogSchema requires a published run to name its own pull request, so this entry could not be written until that number existed. Both commits are in the one pull request and squash to one commit, so the data and the record of why it changed still land together and revert together.
DeployNot runNot run, and recorded as not-run rather than as a pass. This entry is committed into the very pull request whose merge triggers the Pages deploy, so at the moment it was written the deploy could not have happened and no probe available here could have reached it. The confirmation is performed after the merge, with both control arms, and recorded on this run's summary issue, which is where a reader should check it. Writing "ran" here would have been a claim about the future dressed as a measurement.

What was found

Scouts: 44
Pages fetched and hashed: 296
Claims proposed: 441

Claims per creator bundle, with the review threshold its profile set
Creator	Policy	Threshold	Claims
01-ai	long-tail	3-of-3	6
ai-singapore	long-tail	3-of-3	4
ai2	long-tail	3-of-3	8
ai21-labs	long-tail	3-of-3	6
aleph-alpha	long-tail	3-of-3	5
alibaba-cloud	pilot	2-of-3	17
amazon	pilot	2-of-3	14
anthropic	pilot	2-of-3	32
apple	long-tail	3-of-3	9
baidu	long-tail	3-of-3	5
bytedance-seed	long-tail	3-of-3	5
cohere	long-tail	3-of-3	10
cross-creator	long-tail	3-of-3	23
databricks	long-tail	3-of-3	4
deepseek	long-tail	3-of-3	10
eleutherai	long-tail	3-of-3	8
google-deepmind	pilot	2-of-3	7
hugging-face	long-tail	3-of-3	28
ibm	long-tail	3-of-3	15
kyutai	long-tail	3-of-3	4
lelapa-ai	long-tail	3-of-3	4
lg-ai-research	long-tail	3-of-3	7
liquid-ai	long-tail	3-of-3	5
maritaca-ai	long-tail	3-of-3	2
meta	pilot	2-of-3	32
microsoft	pilot	2-of-3	21
minimax	long-tail	3-of-3	10
mistral-ai	long-tail	3-of-3	16
moonshot-ai	long-tail	3-of-3	8
naver	long-tail	3-of-3	5
nous-research	long-tail	3-of-3	1
nvidia	long-tail	3-of-3	12
openai	pilot	2-of-3	20
openbmb	long-tail	3-of-3	12
reka-ai	long-tail	3-of-3	5
sakana-ai	long-tail	3-of-3	4
sarvam-ai	long-tail	3-of-3	4
snowflake	long-tail	3-of-3	4
stability-ai	long-tail	3-of-3	9
tencent	long-tail	3-of-3	9
tii	long-tail	3-of-3	8
upstage	long-tail	3-of-3	4
xai	long-tail	3-of-3	8
xiaomi	long-tail	3-of-3	5
zhipu-ai	long-tail	3-of-3	6

What those claims proposed to do
Kind	Count	Effect
Change	441	Every claim this run proposed is a re-verification of a record that already exists: a lastCheckedDate or verifiedAt moved to 2026-09-24 because the cited page was fetched again and still says what the dataset records. Nothing was added, nothing was removed, and no model fact changed value anywhere in the dataset.

Not covered

A creator that was scouted and produced no claim is not represented here as an absence of scouting. All 44 creators produced a bundle, so unswept is empty; a zero-claim bundle would mean the channels were read and said nothing new.
The sweep reads the primary channels catalogued in sources.json. A creator that announced a model somewhere this dataset does not yet catalogue would not be seen by it, and the discovery pass that looks for exactly that is name-based and would miss a release announced under no recognisable name.

Degraded discovery channels

openai — official-announcement: https://openai.com/index/gpt-4-1/ returned http-403
openai — official-announcement: https://openai.com/index/gpt-5-6/ returned http-403
openai — official-announcement: https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/ returned http-403
openai — official-announcement: https://openai.com/index/previewing-gpt-5-6-sol/ returned http-403
openai — official-announcement: https://openai.com/index/introducing-gpt-5-5/ returned http-403
openai — official-announcement: https://openai.com/index/introducing-gpt-5-4/ returned http-403
alphabet — official-docs: https://abc.xyz/ returned http-403

What was evaluated

Reviewers: 3
Verdicts cast: 1,323
Accepted by panel: 437
Rejected by panel: 4

Deterministic gates and required checks — 8 of 9 checks passed. Exit 0 is a pass; exit 2 means the gate could not run and is never treated as one.
Check	Scope	Exit	Result
`gate-source-approval`	all 45 claim bundles, run before any claim was applied	0	PassAll 45 bundles passed against anchor 64f70ef9. 273 distinct inherited sources and zero proposed sources: this run cites only origins already approved in the catalogued profiles, so there is no case here of a run introducing a source and citing it in the same breath.
`gate-evidence`	the 397 applied claims, in the 43 bundles it admitted, run before any claim was applied	0	PassEvery claim carries a verbatim quote of at least the minimum length, a sha256 content hash, an https credential-free URL, a retrieval of "fetch" and a non-future fetch date, and every claim met the threshold its bundle policy sets -- 2-of-3 for the seven reviewed profiles, 3-of-3 for the long-tail. The threshold was derived by the gate from the reviewed-profile set on disk, not read from the bundles.
`gate-evidence`not required	the 32 claims in the anthropic bundle and the 12 in the nvidia bundle	1	FailExit 1 on both. It refused the anthropic bundle over three claims that reached 1 of the 2 accepts the pilot policy requires, and the nvidia bundle over families--nvidia-cosmos--verifiedAt reaching 2 of the 3 the long-tail policy requires. This gate refuses at bundle granularity, so those four claims took 40 unanimously accepted siblings with them. The run honoured that refusal exactly: nothing from either bundle is in this pull request, which is why this gate run governs no published record and is marked not-required. Not-required records that the refusal was obeyed, never that it was waived -- the 40 siblings are in withheld[] under dropped-after-acceptance rather than quietly applied.
`gate-dataset`	the whole dataset after the 397 edits were applied	0	PassPassed with zero failures. It read requiredCollections and lifecycleStatus out of the schema rather than a hand-copied vocabulary, reporting sources, organizations, families and releases as the floored collections.
`gate-scope`	the branch diff against its computed merge-base	0	PassThe change is confined to the ADR 0003 qualifying class: four dataset documents plus this ledger. Nothing outside web/src/data/ is touched, which is what makes this change eligible to merge unattended at all.
`gate-ledger`	this entry against the diff it describes	0	PassReconciled rather than transcribed: the branch changes dataset documents, so the record counts in posted.documents were re-derived at the anchor and in the working tree and checked against what this entry claims, in both directions. transcription is false, which is the reading that says these numbers were actually checked.
`gate-reversals`	refresh-runs.json and the dataset read together	0	PassNo rejected-by-panel record sits in the dataset unannotated as a result of this run. The four claims this panel refused are field re-verifications of records that were already present, which the gate correctly does not treat as reversals: the records stay, and only their date fields failed to move.
`npm run validate`	the whole web workspace at the dataset commit	0	Pass139 test files, 3,189 of 3,189 tests passing, and astro check reporting 0 errors, 0 warnings and 0 hints across 287 files. Read from an unpiped invocation: an earlier run of the same command through a Select-String filter reported exit 1 while every test passed, which is the shell terminating the producer rather than a failure, and reading that as a result would have manufactured a defect that did not exist.
`ci-preflight`not required	the pull-request checks this branch's diff actually triggers	0	PassRun from the repository root before hand-off, selecting checks from the merge-base diff. It prints its own uncovered set on every run, including the networked link-health sweep, so a green preflight is not read here as a green CI. Not-required because it is a local selector for CI rather than a check CI runs.

Applied over a recorded dissent

These met their threshold and were applied. The objection stands on the record and was not overruled.

families--anthropic-claude-5--verifiedAtEditorial and entity boundariesThe Claude 5 family record is being advanced with a quoted passage about the server-side fallbacks API parameter and Anthropic's recommended models. API fallback guidance is an API-surface/product behavior, not a Claude 5 family boundary statement, so the verification blurs the family/API boundary.
releases--anthropic-claude-fable-5--verifiedAtEditorial and entity boundariesThe target is the legacy Claude Fable 5 release, but the overview and API-release-note passages used here discuss Claude Fable 5.1 and other sibling/successor guidance. Fable 5 and Fable 5.1 are distinct releases in the current record, so a Fable 5.1 passage cannot editorially re-verify Fable 5.
releases--anthropic-claude-opus-5--verifiedAtEditorial and entity boundariesThe target is Claude Opus 5, but one cited overview passage recommends Claude Opus 5.5 for most workloads and discusses Claude Fable 5.1. Opus 5.5 and Fable 5.1 are separate releases, so using that sibling/successor guidance to advance Opus 5 crosses release boundaries.

Posted 397 edits

397 edits across 4 documents, a net change of 0 records.

Dataset documents this run changed
Document	Before	After	What changed
`web/src/data/sources.json`	296	296	248 field-level re-verifications moved to 2026-09-24. No record was added or removed, which is why the counts are equal at both ends; every changed line is a lastCheckedDate or verifiedAt value.
`web/src/data/organizations.json`	44	44	29 field-level re-verifications moved to 2026-09-24. No record was added or removed, which is why the counts are equal at both ends; every changed line is a lastCheckedDate or verifiedAt value.
`web/src/data/families.json`	87	87	64 field-level re-verifications moved to 2026-09-24. No record was added or removed, which is why the counts are equal at both ends; every changed line is a lastCheckedDate or verifiedAt value.
`web/src/data/releases.json`	129	129	56 field-level re-verifications moved to 2026-09-24. No record was added or removed, which is why the counts are equal at both ends; every changed line is a lastCheckedDate or verifiedAt value.

Each document links to the file as this run left it, not as it stands today.

Not posted 46 items

Rejected by the review panel

sources--anthropic-sonnet-4-5-announcement--lastCheckedDateReached 1 of the 2-of-3 its pilot policy requires, so the panel refused it. provenance: The proposed lastCheckedDate is for the source record named ?Introducing Claude Sonnet 4.5,? but the only quote says ?We?re announcing a new life sciences research group and laboratory at Anthropic? and describes Claude discovering an enzyme system. Read alone, that quote does not state or identify Claude Sonnet 4.5 or the named announcement page. editorial: The source record is named as 'Introducing Claude Sonnet 4.5', but the quoted passage is about Anthropic's life-sciences research group and Claude discovering an enzyme system. Treating that research/product news passage as the Sonnet 4.5 model-release source crosses the product/research versus model-release boundary.Blocked by reviewer:provenancereviewer:editorial
sources--anthropic-sonnet-4-6-announcement--lastCheckedDateReached 1 of the 2-of-3 its pilot policy requires, so the panel refused it. provenance: The proposed lastCheckedDate is for the source record named ?Introducing Claude Sonnet 4.6,? but the quote says ?We?re announcing a new life sciences research group and laboratory at Anthropic? and describes a CRISPR-like enzyme result. The quote does not mention Claude Sonnet 4.6 or otherwise state that the fetched page is still the source record named by the dataset. editorial: The source record is named as 'Introducing Claude Sonnet 4.6', but the quoted passage is the same Anthropic life-sciences/Claude enzyme-system post rather than a Sonnet 4.6 release statement. That makes the source identity a product/research item, not the model-release page the record names.Blocked by reviewer:provenancereviewer:editorial
families--anthropic-claude-4-6--verifiedAtReached 1 of the 2-of-3 its pilot policy requires, so the panel refused it. provenance: This verifiedAt claim depends on all three cited sources, but the attached quote for anthropic-sonnet-4-6-announcement is the life-sciences-laboratory passage, not a Claude Sonnet 4.6 or Claude 4.6 passage. Because one cited source quote does not say the record-specific fact it is being used to re-verify, the proposed 2026-09-24 verifiedAt is not supported by the claim?s own evidence. editorial: The Claude 4.6 family evidence set includes the supposed Sonnet 4.6 announcement source, but the quoted passage is the Anthropic life-sciences/Claude enzyme-system post. Using that research/product news item to re-verify a Claude 4.6 model family crosses entity boundaries.Blocked by reviewer:provenancereviewer:editorial
families--nvidia-cosmos--verifiedAtReached 2 of the 3-of-3 its long-tail policy requires, so the panel refused it. editorial: Reject on entity boundary: the NVIDIA profile explicitly warns that Cosmos is a World Foundation Model Platform and that platform names are not family names. Re-verifying a `families` record named Cosmos would bless a platform-as-family boundary; the quotes refer to Cosmos broadly or to an agreement's 'NVIDIA Cosmos Model', not to a ModelTree family.Blocked by reviewer:editorial

Accepted by the panel, then dropped

sources--anthropic-models-overview--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-fable-5-mythos-5-docs--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-fable-5-1-announcement--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-fable-5-1-docs--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-fable-5-docs--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-model-deprecations--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-opus-5-announcement--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-haiku-4-5-announcement--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-sonnet-5-docs--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-sonnet-5-announcement--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-opus-4-8-announcement--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-opus-4-7-announcement--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-opus-4-6-announcement--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-opus-4-8-docs--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-mythos-5-1-docs--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-api-release-notes--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--anthropic-opus-5-docs--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
organizations--anthropic--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
families--anthropic-claude-5--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
families--anthropic-claude-4-7--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
families--anthropic-claude-4-8--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
families--anthropic-claude-5-1--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
releases--anthropic-claude-fable-5--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
releases--anthropic-claude-opus-5--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
releases--anthropic-claude-opus-4-6--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
releases--anthropic-claude-opus-4-7--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
releases--anthropic-claude-opus-4-8--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
releases--anthropic-claude-fable-5-1--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
releases--anthropic-claude-mythos-5-1--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole anthropic bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:anthropic
sources--nvidia-nemotron-4-340b-announcement--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole nvidia bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:nvidia
sources--nvidia-nemotron-4-340b-research--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole nvidia bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:nvidia
sources--nvidia-nemotron-4-340b-model-card--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole nvidia bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:nvidia
sources--nvidia-nemotron-nano-9b-v2-model-card--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole nvidia bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:nvidia
sources--nvidia-cosmos-announcement--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole nvidia bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:nvidia
sources--nvidia-cosmos-diffusion-7b-model-card--lastCheckedDateAccepted by the panel, then withheld because gate-evidence refused the whole nvidia bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:nvidia
organizations--nvidia--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole nvidia bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:nvidia
families--nvidia-nemotron-4-340b--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole nvidia bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:nvidia
families--nvidia-nemotron-nano-2--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole nvidia bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:nvidia
releases--nvidia-nemotron-4-340b-base--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole nvidia bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:nvidia
releases--nvidia-nemotron-nano-9b-v2--verifiedAtAccepted by the panel, then withheld because gate-evidence refused the whole nvidia bundle over a sibling claim that fell below threshold. That gate refuses at bundle granularity; applying this claim alone would have been cherry-picking past a refusal, so it was obeyed rather than worked around.Blocked by gate-evidencebundle:nvidia

Blocked by policy before it could run

anthropic-claude-opus-5-5Still not proposed. route:home measured 1,102,773 bytes against its 1,105,000 ceiling, leaving 2,227 bytes of headroom against a measured 7,002 bytes per additional creator, so no creator fits. Re-measured this run rather than recalled; the shortfall is tracked on issue #1172.Blocked by web/asset-budgets.json:route:homenpm run assets:runway
google-deepmind-gemini-3-8-flashWithheld on the same measurement as Claude Opus 5.5: route:home has 2,227 bytes spare and each additional creator costs a measured 7,002 bytes. Third consecutive run in which the budget, not the evidence, is what holds these back.Blocked by web/asset-budgets.json:route:homenpm run assets:runway

What this run does not prove

Every claim this run proposed is a date move. A green run therefore re-verifies that a cited page was fetched and still carries a passage naming the record; it does not establish that any model fact changed, because none did.
Fourteen of the 373 distinct quotes this run extracted are bound to more than one source page, touching 35 claims across 25 pages. Most are legitimate -- a repository README and a model card genuinely sharing a passage -- but the anthropic-sonnet-4-5 and anthropic-sonnet-4-6 pair is the case where it signals a site-wide promotional component being mistaken for article body, and it is what the panel caught. The measurement is the chair's and is published as a finding; it was not used to overrule any reviewer, and the other 12 shared quotes were left to the panel to judge on their own terms.
Seven catalogued channels returned HTTP 403 and were not read this run: six openai.com/index pages and abc.xyz. Those records were re-verified from the other sources cited on them where any existed, and where none did the claim was excluded rather than asserted. No user-agent was spoofed to get past a refusal.
gate-evidence checks the form of evidence -- that a quote exists, is long enough, is verbatim in the fetched body, and carries a hash and a date. It does not re-fetch the page or check that the hash still matches what the origin serves today, which ADR 0005 states plainly. The hashes here are of bytes this run received on 2026-09-24 and are checkable against the saved bodies, not against the live web.
115 candidate claims were excluded before review, 73 of them because the cited source returned a body that contained no substantive passage naming the record. An exclusion is a claim not made, which is a weaker statement than a claim examined and rejected, and the two are not interchangeable.
The 40 claims withheld as dropped-after-acceptance were each accepted unanimously or at threshold by the panel. They are absent from this pull request because of where they sat, not because of what the reviewers thought of them, and they will be proposed again by the next run.
A date-only change moved a measured byte, and the run could not say which one. route:home measured 1,102,774 raw bytes with this run's data against 1,102,773 with the dataset restored to the anchor -- same command, same session, same installed dependencies, four readings, two each way, perfectly correlated with the data state. Every applied edit replaces a ten-character date with another ten-character date, so no input to that measurement changed length, and the figure it moved is a raw byte count with no compressor in it. The mechanism was not identified. An attempt to locate it by diffing dist/ was discarded rather than reported: dist/ was 40 minutes stale throughout, because assets:runway builds its own two arms and never writes dist/, so those readings were taken against the wrong tree and say nothing. What the one byte does not do is change a decision -- both arms read REFUSED at 0 creators affordable, so the finding tracked on issue #1172 is the same either way, and one byte against the 2% drift allowance fails nothing.

Follow-ups — proposed, not fixed

The anthropic-sonnet-4-5 and anthropic-sonnet-4-6 announcement URLs both returned 200 with no redirect and different byte counts, yet yielded an identical passage. Whether those two records still cite the right URLs is a question for a reviewed change, not for an unattended refresh, and until it is answered those two sources will keep failing this panel.
The NVIDIA Cosmos record is held as a family while NVIDIA's own profile warns that Cosmos is a World Foundation Model Platform. That is an entity-boundary question the refresh loop cannot settle on its own, because resolving it means moving or renaming a record rather than re-verifying one.
route:home has 2,226 bytes of headroom against a measured 7,001 bytes per additional creator, so no creator fits and two releases stay withheld for a third consecutive run. Tracked on issue #1172.
Find what makes a raw route measurement move by one byte when every input to it changed value without changing length. This run reproduced the move four times and did not explain it, and a byte that nobody can account for is worth more attention than a byte that is simply large: the same unexplained mechanism at a larger amplitude would be read as ordinary growth. Locating it means instrumenting the two arms assets:runway builds internally, not the dist/ tree, which that command does not write.

2026-09-23-02b51c

Verification dates refreshed across 45 bundles; five claims rejected and two new releases withheld on the asset budget

Scope requested: Dataset JSON under web/src/data/ only, per ADR 0003. Four documents changed — sources, organizations, families and releases — plus this ledger. gate-scope reported nothing outside the qualifying class.PublishedSep 23, 2026430 edits posted · 2 items withheld

A full-catalogue pass: 44 creator bundles plus one discovery bundle scouted from primary sources, 296 catalogued pages fetched and content-hashed in a single English-preferring sweep, 435 claims put to a three-rubric panel, and 430 applied. Every applied claim advances a verification date in place against freshly fetched evidence, so the dataset gained no record and lost none. Five claims were rejected by the panel. Two genuinely new releases — Claude Opus 5.5 and Gemini 3.8 Flash — were fully evidenced and deliberately withheld because either one alone breaches the home page critical-asset ceiling. The panel found two real defects in this run’s own generator before any of it reached the dataset.

PreflightRanTrunk anchor resolved to d10050270248b4c34f198cd357d99e2c3d227ebe, reflog-dated 2026-09-22 08:21:23 -0400. Working tree clean and HEAD level with trunk at the start of the run. npm was probed both ways: the bare name is refused by the PowerShell execution policy — installed and blocked, not absent — while npm.cmd reports 11.9.0, and that is the form every npm command here used. The execution policy was not changed. npm ci ran at exit 0 with the lockfile provably untouched (ADR 0004).
ScoutRan296 catalogued sources fetched and content-hashed; 281 verified and 15 came back degraded, carrying no claim. The fetcher was given an explicit accept-language of en-US after Google content-negotiated Chinese and Polish bodies mid-run; the entire base was then re-fetched through one code path at one instant so no quote cites a body obtained under different headers. The marker matcher was controlled two-sided at 14 arms — 5 positive, 9 negative, 0 failures — including the prefix-boundary arm that caught last run's GPT-Image-2 defect. Discovery surfaced exactly two genuinely new releases, anthropic/claude-opus-5-5 and google/gemini-3-8-flash, both attested on already-trusted origins.
ReviewRan435 claims judged by three independent rubrics — provenance, consistency and editorial — dispatched in parallel with no shared context and no reviewer shown another's verdict. Claims were batched per reviewer, never across. The planted-defect control was read before the real tally: 8 of 8, six synthetic defects rejected unanimously by all three rubrics and two ordinary claims accepted, so the arms differ and the panel measures something. The panel then caught two real defects in this run's generator, described in the caveats. Six claims changed as a result and were re-judged by a fresh panel; its must-reject arm passed 3-of-3 and its must-accept arm was void for the consistency rubric — the synthetic target id is absent from the dataset, which that rubric correctly refuses — so a supplementary two-sided consistency control was run on two real records and passed 2 of 2 with the arms differing.
GatesRangate-source-approval 45/45 at exit 0, two-sided against a fabricated unapproved-origin bundle at exit 1. gate-evidence 45/45 at exit 0 with a three-arm control in the same session — untouched bundle exit 0, a sub-threshold change exit 1, an evidence-stripped claim exit 1 — so the pass is a reading and not a constant. Before the five rejected claims were demoted, gate-evidence independently re-derived the same five sub-threshold claims at the same policies, which is corroboration from an instrument that never saw the tally. gate-dataset exit 0. gate-reversals exit 0. gate-scope and gate-ledger reported below. Every exit code was read from an unpiped invocation: an earlier piped read of gate-reversals returned -1, the PowerShell early-termination value, and was discarded rather than believed.
PublishRan430 claims applied from a clean tree, 0 refused and 0 skipped, as exactly 430 insertions and 430 deletions across four documents — byte-neutral, as a pure date advance must be. The applier carries a hard precondition that it must reproduce a file byte-for-byte before rewriting it; that precondition was added after a round-trip control showed publishers.json cannot be reproduced by JSON.stringify, and it would have refused rather than reformat 200 lines of out-of-scope content. The diff was then asserted independently: 430 date pairs, 430 forward, 0 backward, 0 unchanged, 0 non-date, a single distinct new value of 2026-09-23, and the ledger append-only with the prior 28 entries byte-identical. No force, no admin override, no skipped gate, no lowered threshold and no direct push to main.
DeployNot runNot run at the moment this entry was written, and recorded as not-run rather than as a pass: the entry is committed before the merge, so no GitHub Pages deployment for the merge commit can exist to be checked. A check that was not run is not a check that passed. The verification — the returned headSha compared against the full 40-character merge SHA, with a positive control and an all-zeros negative arm required to disagree before either is read — is performed after the merge and reported in the run summary issue.

What was found

Scouts: 45
Pages fetched and hashed: 296
Claims proposed: 435

Claims per creator bundle, with the review threshold its profile set
Creator	Policy	Threshold	Claims
openai	pilot	2-of-3	32
anthropic	pilot	2-of-3	33
google-deepmind	pilot	2-of-3	10
meta	pilot	2-of-3	32
xai	long-tail	3-of-3	11
mistral-ai	long-tail	3-of-3	14
deepseek	long-tail	3-of-3	7
alibaba-cloud	pilot	2-of-3	13
microsoft	pilot	2-of-3	21
amazon	pilot	2-of-3	14
cohere	long-tail	3-of-3	10
ai2	long-tail	3-of-3	7
tii	long-tail	3-of-3	7
nvidia	long-tail	3-of-3	10
ai21-labs	long-tail	3-of-3	8
zhipu-ai	long-tail	3-of-3	5
moonshot-ai	long-tail	3-of-3	6
eleutherai	long-tail	3-of-3	8
lg-ai-research	long-tail	3-of-3	7
snowflake	long-tail	3-of-3	4
upstage	long-tail	3-of-3	3
ibm	long-tail	3-of-3	12
baidu	long-tail	3-of-3	4
tencent	long-tail	3-of-3	7
bytedance-seed	long-tail	3-of-3	5
stability-ai	long-tail	3-of-3	8
databricks	long-tail	3-of-3	4
minimax	long-tail	3-of-3	9
apple	long-tail	3-of-3	7
hugging-face	long-tail	3-of-3	31
01-ai	long-tail	3-of-3	6
sakana-ai	long-tail	3-of-3	5
sarvam-ai	long-tail	3-of-3	4
naver	long-tail	3-of-3	5
aleph-alpha	long-tail	3-of-3	4
reka-ai	long-tail	3-of-3	4
nous-research	long-tail	3-of-3	4
liquid-ai	long-tail	3-of-3	4
xiaomi	long-tail	3-of-3	4
ai-singapore	long-tail	3-of-3	4
kyutai	long-tail	3-of-3	4
lelapa-ai	long-tail	3-of-3	3
maritaca-ai	long-tail	3-of-3	5
openbmb	long-tail	3-of-3	9
discovery	long-tail	3-of-3	21

What those claims proposed to do
Kind	Count	Effect
Change	430	Advances a verification date in place against freshly fetched evidence; adds no record.
Unchanged	5	Re-checked and not applied. These are the five claims the panel rejected, demoted from "change" so that nothing sub-threshold is marked as changing data, with statement, evidence and all three verdicts kept verbatim.

Not covered

Creator sites that publish no machine-readable model catalogue are swept only for the pages already in the approved source list; a model announced solely in a blog post outside that list is not discovered.
The four opensource.org channels began returning HTTP 403 partway through this run, so no osiApproved field was touched anywhere and 80 records could not have their verification date advanced.
Serving-platform catalogues are deliberately out of scope: a platform offering a model states nothing about how its creator releases it.
Four releases are pinned by model-fit statements dated 2026-08-31 and were excluded from verification-date advancement rather than having those statements re-dated, which would have been an editorial judgement this run declined to make unattended.

Degraded discovery channels

openai — https://openai.com/index/gpt-4-1/: Fetch returned HTTP 403, so no usable evidence was obtained this run. No claim rests on this channel; the records citing it keep their prior verification date rather than being advanced on stale evidence. The "creator" field here names the owning organization, not the source id — the 2026-09-22 entry recorded the source id in this field and that is corrected from this run on.
openai — https://openai.com/index/gpt-5-6/: Fetch returned HTTP 403, so no usable evidence was obtained this run. No claim rests on this channel; the records citing it keep their prior verification date rather than being advanced on stale evidence. The "creator" field here names the owning organization, not the source id — the 2026-09-22 entry recorded the source id in this field and that is corrected from this run on.
openai — https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/: Fetch returned HTTP 403, so no usable evidence was obtained this run. No claim rests on this channel; the records citing it keep their prior verification date rather than being advanced on stale evidence. The "creator" field here names the owning organization, not the source id — the 2026-09-22 entry recorded the source id in this field and that is corrected from this run on.
openai — https://openai.com/index/previewing-gpt-5-6-sol/: Fetch returned HTTP 403, so no usable evidence was obtained this run. No claim rests on this channel; the records citing it keep their prior verification date rather than being advanced on stale evidence. The "creator" field here names the owning organization, not the source id — the 2026-09-22 entry recorded the source id in this field and that is corrected from this run on.
openai — https://openai.com/index/introducing-gpt-5-5/: Fetch returned HTTP 403, so no usable evidence was obtained this run. No claim rests on this channel; the records citing it keep their prior verification date rather than being advanced on stale evidence. The "creator" field here names the owning organization, not the source id — the 2026-09-22 entry recorded the source id in this field and that is corrected from this run on.
openai — https://openai.com/index/introducing-gpt-5-4/: Fetch returned HTTP 403, so no usable evidence was obtained this run. No claim rests on this channel; the records citing it keep their prior verification date rather than being advanced on stale evidence. The "creator" field here names the owning organization, not the source id — the 2026-09-22 entry recorded the source id in this field and that is corrected from this run on.
google-deepmind — https://storage.googleapis.com/deepmind-media/gemini/gemini_v2_5_report.pdf: The channel serves a PDF, which this fetcher reads as bytes rather than as quotable text. No claim rests on this channel; the records citing it keep their prior verification date rather than being advanced on stale evidence. The "creator" field here names the owning organization, not the source id — the 2026-09-22 entry recorded the source id in this field and that is corrected from this run on.
(no creator: publisher alphabet maps to no organization) — https://abc.xyz/: Fetch returned HTTP 403, so no usable evidence was obtained this run. No claim rests on this channel; the records citing it keep their prior verification date rather than being advanced on stale evidence. The "creator" field here names the owning organization, not the source id — the 2026-09-22 entry recorded the source id in this field and that is corrected from this run on.
openai — https://developers.openai.com/api/docs/models: The page fetched cleanly but no longer contains the marker the catalogue records for it. No claim rests on this channel; the records citing it keep their prior verification date rather than being advanced on stale evidence. The "creator" field here names the owning organization, not the source id — the 2026-09-22 entry recorded the source id in this field and that is corrected from this run on.
(no creator: publisher open-source-initiative maps to no organization) — https://opensource.org/license/apache-2.0: Fetch returned HTTP 403, so no usable evidence was obtained this run. No claim rests on this channel; the records citing it keep their prior verification date rather than being advanced on stale evidence. The "creator" field here names the owning organization, not the source id — the 2026-09-22 entry recorded the source id in this field and that is corrected from this run on.
(no creator: publisher open-source-initiative maps to no organization) — https://opensource.org/license/mit: Fetch returned HTTP 403, so no usable evidence was obtained this run. No claim rests on this channel; the records citing it keep their prior verification date rather than being advanced on stale evidence. The "creator" field here names the owning organization, not the source id — the 2026-09-22 entry recorded the source id in this field and that is corrected from this run on.
(no creator: publisher open-source-initiative maps to no organization) — https://opensource.org/licenses: Fetch returned HTTP 403, so no usable evidence was obtained this run. No claim rests on this channel; the records citing it keep their prior verification date rather than being advanced on stale evidence. The "creator" field here names the owning organization, not the source id — the 2026-09-22 entry recorded the source id in this field and that is corrected from this run on.
zhipu-ai — https://z.ai/blog/glm-4.5: The page fetched at HTTP 200 but returned a body too thin to quote, in both sweeps. No claim rests on this channel; the records citing it keep their prior verification date rather than being advanced on stale evidence. The "creator" field here names the owning organization, not the source id — the 2026-09-22 entry recorded the source id in this field and that is corrected from this run on.
apple — https://huggingface.co/apple/FastVLM-7B/raw/main/LICENSE: The page fetched cleanly but no longer contains the marker the catalogue records for it. No claim rests on this channel; the records citing it keep their prior verification date rather than being advanced on stale evidence. The "creator" field here names the owning organization, not the source id — the 2026-09-22 entry recorded the source id in this field and that is corrected from this run on.
(no creator: publisher open-source-initiative maps to no organization) — https://opensource.org/ai/open-source-ai-definition: Fetch returned HTTP 403, so no usable evidence was obtained this run. No claim rests on this channel; the records citing it keep their prior verification date rather than being advanced on stale evidence. The "creator" field here names the owning organization, not the source id — the 2026-09-22 entry recorded the source id in this field and that is corrected from this run on.

What was evaluated

Reviewers: 3
Verdicts cast: 1,355
Accepted by panel: 430
Rejected by panel: 5

Deterministic gates and required checks — 9 of 13 checks passed. Exit 0 is a pass; exit 2 means the gate could not run and is never treated as one.
Check	Scope	Exit	Result
`gate-source-approval`	all 45 bundles	0	PassEvery cited origin is in the approved list at trunk d100502702. Invoked once per bundle, 45 invocations, all exit 0.
`gate-source-approval`not required	fabricated unapproved origin, control arm	1	FailThe failing arm that makes the passing arm readable. Built deliberately because every real claim this run rests on an already-trusted origin, so without it the gate would have been one-sided.
`gate-evidence`	all 45 bundles, after demotion	0	PassEvery applied claim carries a url, a verbatim quote, a content hash and a fetch date, and meets the threshold its profile policy requires. The gate derives that policy from tools/updater/profiles itself and never from the bundle, so its agreement with the panel tally is independent corroboration.
`gate-evidence`not required	all 45 bundles, before demotion	1	FailThe first sweep refused 5 of 45 bundles, naming exactly the five claims the panel had rejected, at exactly the policies the panel applied — an instrument that never saw the tally re-deriving it. Those five were then demoted from "change" to "unchanged" with statement, evidence and all three verdicts kept verbatim, because nothing was applied for them.
`gate-evidence`not required	three-arm control on one real bundle	1	FailRun in the same session against the same binary: the untouched bundle exited 0, the same bundle with one accepted change forced sub-threshold exited 1, and the same bundle with one claim’s evidence emptied exited 1. The arms differ, so the 45/45 pass is a reading and not a constant.
`gate-dataset`	web/src/data	0	PassAll structural gates passed over 296 sources, 52 publishers, 44 organizations, 87 families and 129 releases. It does not read the ledger schema; npm run validate caught four schema defects in this very entry that gate-dataset passed over, which is recorded here rather than quietly repaired.
`gate-reversals`	ledger, 282 recorded rejections	0	PassNewly load-bearing this run because the panel rejected five claims. No panel-rejected claim is sitting in the dataset unexplained.
`gate-ledger`	branch mode against the trunk anchor	0	Passtranscription false — this branch changes the dataset, so the entry is reconciled against a real diff. posted.documents matches the four changed documents in both directions, record counts were counted at both ends rather than taken from the entry, and the prior 28 entries are untouched.
`gate-ledger`	history mode	0	PassEvery run id declared in a commit subject reachable from the anchor has a matching entry.
`gate-scope`	working tree against the merge-base it computes itself	0	PassFive documents changed, all in class, nothing out of class, and empty false — so the pass is a measurement over a real diff rather than a pass over an empty one.
`gate-scope`not required	out-of-class file present, control arm	1	FailA file was created outside web/src/data and the same gate exited 1 naming it, then the file was removed and the gate returned to exit 0 in the same session. The arms differ, so "nothing out of class" is a reading.
`npm run validate`	web/, committed tree	—	PassReported after this entry was repaired. Its first run over this entry failed with four schema violations — a missing evaluated.gates, two invented withheld categories and an empty references list — every one a real defect in what had been written. They were fixed rather than worked around, and the failure is left recorded here because a validator that caught something is evidence the validator works.
`ci-preflight`	repository root, checks this diff triggers	—	PassSelects the pull-request checks the branch diff actually triggers and runs them locally. Reported after this entry was repaired.

Applied over a recorded dissent

These met their threshold and were applied. The objection stands on the record and was not overruled.

releases:anthropic-claude-fable-5--verifiedAtProvenanceThe quotes do not substantively support the Claude Fable 5 release record: the overview quote names 'claude-fable-5-1' rather than the target, and the Fable pages are title/navigation snippets such as 'Introducing Claude Fable 5 and Claude Mythos 5 ... capabilities, API changes' without the recorded release facts.
releases:anthropic-claude-fable-5--verifiedAtEditorial and entity boundariesOne cited evidence item for this Claude Fable 5 release is an API release-notes quote about Claude Fable 5.1/Claude Mythos 5.1 availability on Google Cloud. That is a successor/platform-serving statement, not a statement about the Claude Fable 5 release itself, so recertifying the target release from it crosses release and serving-platform boundaries.
families:apple-openelm--verifiedAtProvenanceThe apple-openelm-3b-instruct-model-card quote is only page title/navigation ('apple/OpenELM-3B-Instruct · Hugging Face ... Models Datasets Spaces ...') and does not state the OpenELM family facts; navigation chrome supports no re-verification of the record.
families:deepseek-v4--verifiedAtProvenanceThe two model-card quotes are only Hugging Face title/navigation strings ('deepseek-ai/DeepSeek-V4-Pro · Hugging Face ...' and 'deepseek-ai/DeepSeek-V4-Flash · Hugging Face ...') rather than statements about the DeepSeek-V4 family, so not every cited source substantively supports the record.
releases:google-gemini-3-6-flash--verifiedAtProvenanceSeveral cited quotes are only page titles/navigation, and the deprecations quote lists unrelated Gemini 3.1 entries rather than Gemini 3.6 Flash. Those quotes do not state the target release facts, so the evidence does not show every cited source still supports the record.
releases:google-gemini-3-6-flash--verifiedAtEditorial and entity boundariesThe evidence set includes Google Cloud Gemini Enterprise Agent Platform documentation as support for a Google DeepMind Gemini 3.6 Flash release. That page is a serving-platform/product surface, while the rubric requires the Gemini model release to stay distinct from platform offerings, so advancing verifiedAt would overstate what that platform quote establishes about the release.
families:xai-grok-4--verifiedAtProvenanceThe xai-models-docs evidence is only documentation chrome ('Grok Models & Pricing | SpaceXAI Docs API ... Status Loading... Resources ...') and states no Grok 4 family fact, so that cited source does not support the re-verification claim.
organizations:meta--verifiedAtProvenanceThe quoted evidence is mostly page titles/chrome; for example, 'The Llama 4 herd: The beginning of a new era...' and 'Introducing Muse Spark...' name pages but do not state the organization record's claims such as Llama weights being downloadable, Muse being served only through Meta surfaces, or no Muse weights being released. [This claim was accepted: it reached 2 of the 2 accepts its pilot policy requires. The dissent is recorded because the threshold is the rule, not because the objection was found wrong.]

Posted 430 edits

430 edits across 4 documents, a net change of 0 records.

Dataset documents this run changed
Document	Before	After	What changed
`web/src/data/sources.json`	296	296	Every edit advances a verification date in place, so the record count is unchanged by construction.
`web/src/data/organizations.json`	44	44	Every edit advances a verification date in place, so the record count is unchanged by construction.
`web/src/data/families.json`	87	87	Every edit advances a verification date in place, so the record count is unchanged by construction.
`web/src/data/releases.json`	129	129	Every edit advances a verification date in place, so the record count is unchanged by construction.

Not posted 2 items

Blocked by policy before it could run

anthropic-claude-opus-5-5--addClaude Opus 5.5, fully evidenced from an already-trusted origin and deliberately not proposed to the panel for application. Its supporting source record was withheld with it, since a source cited by nothing would fail dataset coherence. Measured on this tree: the home page critical raw payload is 1102774 bytes against a 1105000 ceiling, leaving 2226 bytes of headroom. Adding this release alone takes it to 1111473 bytes and adding both new releases takes it to 1120589 — a breach either way, of 6473 and 15589 bytes. Raising a ceiling in web/asset-budgets.json is forbidden by ADR 0015 and is outside the ADR 0003 qualifying class, so the only honest outcome is to withhold. Held, not rejected: the evidence is carried into the run summary issue. This is the second consecutive run blocked this way, which is filed as a follow-up.Blocked by web/asset-budgets.jsondocs/adr/0015-asset-budget-measurements-are-in-class-ceilings-are-not.md
google-gemini-3-8-flash--addGemini 3.8 Flash, fully evidenced from an already-trusted origin and deliberately not proposed to the panel for application. Its supporting source record was withheld with it, since a source cited by nothing would fail dataset coherence. Measured on this tree: the home page critical raw payload is 1102774 bytes against a 1105000 ceiling, leaving 2226 bytes of headroom. Adding this release alone takes it to 1111473 bytes and adding both new releases takes it to 1120589 — a breach either way, of 6473 and 15589 bytes. Raising a ceiling in web/asset-budgets.json is forbidden by ADR 0015 and is outside the ADR 0003 qualifying class, so the only honest outcome is to withhold. Held, not rejected: the evidence is carried into the run summary issue. This is the second consecutive run blocked this way, which is filed as a follow-up.Blocked by web/asset-budgets.jsondocs/adr/0015-asset-budget-measurements-are-in-class-ceilings-are-not.md

What this run does not prove

15 source channels came back degraded and carry no claim this run; the records they back keep their prior verification date rather than being advanced on stale evidence.
The four opensource.org channels newly degraded mid-run: they succeeded in the first sweep and returned HTTP 403 in the second and on spaced retries. Their stale page files were deleted so that no quote can cite a content hash no longer held. Those four alone account for 80 of the 101 records excluded for having an unusable cited source.
Panel catch, colliding claim ids: eight ids in this dataset exist as both a family and a release, so a claim id of the form <targetId>--<field> collides across collections. All three rubrics independently flagged the duplicates. The generator now prefixes the collection. Votes already cast were re-keyed on the stable triple collection|targetId|field under three hard guards — sheet-order alignment, triple uniqueness and full consumption — because the id is bookkeeping and never what a rubric judged.
Panel catch, evidence coverage: a verifiedAt statement asserts that every cited source was re-fetched, but the generator truncated the attached evidence to three sources. The consistency rubric flagged exactly the six claims of 155 where that truncation hid a cited source — perfect recall and precision on a defect the run did not know it had. The truncation was removed and only those six claims changed; the other 429 were byte-identical, so existing verdicts stayed valid on unchanged content.
Three of those six had been accepted at the 2-of-3 pilot bar over a correct dissent before the defect was fixed, which is a live demonstration that the looser threshold can outvote a true finding. The response was to fix the defect and re-judge, never to overrule the panel.
On re-judging with full evidence attached, the provenance rubric rejected all six for a further reason: the newly visible quotes for the supplementary sources are page titles and site navigation rather than substantive statements. That is the same defect class as the chrome-only arm planted in the control, which all three rubrics had already rejected, so the finding is consistent with the instrument’s measured behaviour. Five of the six failed their threshold and were not applied.
The three-rubric panel is three prompts, not three minds: instances of one model family reading one page share a failure mode. The majority buys independence of reasoning, not of training, and a source that misleads all three misleads the panel.
Four releases are pinned by model-fit statements dated 2026-08-31 — meta-llama-4-scout three times, anthropic-claude-mythos-5, anthropic-claude-haiku-4-5 and openai-gpt-5. Their verification dates were left alone rather than re-dating the statements, which would have been an editorial judgement made unattended.
openai-models-catalog fetched cleanly but no longer names GPT-Image-2. That is a genuine finding about the catalogue rather than a fetch failure, and its lastCheckedDate was deliberately not advanced.
Two further Gemini entries, gemini-3.8-live and gemini-3.8-live-extended-thinking, were observed in the fetched pages and not added: neither is evidenced well enough to stand as a release record, and the asset budget could not have absorbed them in any case.

Follow-ups — proposed, not fixed

Second consecutive run blocked by the home page critical-asset ceiling. Two fully-evidenced releases cannot ship because one release costs roughly 8700 critical bytes and only 2226 remain. The ceiling is correct and ADR 0015 correctly forbids widening it from inside a refresh, so the payload needs to shrink or the home page needs to stop carrying the whole dataset — either way it is a change outside the ADR 0003 class and needs its own issue.
The generator picks a quote per cited source by best match and falls back to the head of the document when no substantive match exists, which yields page titles and navigation. The provenance rubric rejects those correctly, so the failure is loud rather than silent, but a fallback that returned no quote at all — forcing the record to be excluded rather than weakly evidenced — would be more honest than one that returns chrome.
The four opensource.org channels should be probed on a schedule independent of a refresh run. Losing them costs 80 record exclusions, and discovering that inside a run means the run absorbs the cost.
Claim ids should carry their collection at the point they are minted rather than being repaired after a panel notices the collision. Eight ids in the current dataset are both a family and a release, so the collision is structural and will recur.

2026-09-22-13c600

Verification dates refreshed across 44 creator bundles; six claims rejected and six withheld

Scope requested: Dataset JSON under web/src/data/ only, per ADR 0003. gate-scope reported four documents changed and nothing outside the qualifying class.PublishedSep 22, 2026340 edits posted · 6 items withheld

A full-catalogue pass: 44 creator bundles plus one discovery bundle scouted from primary sources, 296 pages fetched and hashed, 352 claims put to a three-rubric panel, and 340 applied. Every applied claim advances a verification date in place against freshly fetched evidence, so the dataset gained no record and lost none. Six claims were rejected by the panel and six more were accepted and then deliberately withheld. Six synthetic defects were planted into the review payloads as a control and all six were caught.

PreflightRanTrunk anchor resolved to 5f32f323d270897868cbd82dfe86b16d212644a0, reflog-dated 2026-09-21 08:39:03 -0400. Working tree clean and HEAD equal to trunk at the start of the run. The Drydock CLI was probed both ways: the bare name is blocked by the PowerShell execution policy, while drydock.cmd reports 0.1.0 — installed and runnable, not absent.
ScoutRan296 catalogued sources fetched and content-hashed; 7 channels came back degraded and carry no claim. Discovery over the fetched pages surfaced exactly one genuinely new release, deepseek-ai/DeepSeek-V4.1-Flash.
ReviewRan352 claims judged by three independent rubrics — provenance, consistency and editorial — launched in parallel with no shared context and no reviewer shown another's verdict. 15 reviewer agents returned verdict files; all 15 validated structurally with exact row coverage. The planted-defect control arm was read before the real tally: 6 of 6 caught.
GatesRangate-evidence and gate-source-approval each 44/44 at exit 0 over the applied set, both two-sided: evidence returned exit 1 over the unsplit bundles, and source-approval returned exit 1 against a fabricated unapproved-origin bundle. gate-dataset exit 0 over 645 records. gate-scope exit 0 with four documents changed, so the pass is a measurement and not an empty diff. gate-ledger exit 0 reporting transcription false.
PublishRan340 claims applied from a clean tree, 0 refused and 0 skipped, as exactly 340 insertions and 340 deletions across four documents — byte-neutral, as a pure date advance must be. npm run validate then passed 3189 tests across 139 files on the committed tree. Pull request #1176 was opened from commit e73391e5b3 with auto-merge enabled, squashing once web-ci is green. No force, no admin override, no skipped gate and no direct push to main.
DeployNot runNot yet run at the moment this entry was written, and recorded as not-run rather than as a pass: the entry is committed before the merge, so no GitHub Pages deployment for the merge commit can exist to be checked. A check that was not run is not a check that passed. The verification — the returned headSha compared against the full 40-character merge SHA, with a positive control and a negative all-zeros arm required to disagree before either is read — is performed after the merge and reported in the run summary issue.

What was found

Scouts: 45
Pages fetched and hashed: 296
Claims proposed: 352

Claims per creator bundle, with the review threshold its profile set
Creator	Policy	Threshold	Claims
01-ai	long-tail	3-of-3	6
ai-singapore	long-tail	3-of-3	4
ai2	long-tail	3-of-3	5
ai21-labs	long-tail	3-of-3	8
aleph-alpha	long-tail	3-of-3	4
alibaba-cloud	pilot	2-of-3	13
amazon	pilot	2-of-3	14
anthropic	pilot	2-of-3	26
apple	long-tail	3-of-3	7
baidu	long-tail	3-of-3	4
bytedance-seed	long-tail	3-of-3	5
cohere	long-tail	3-of-3	10
databricks	long-tail	3-of-3	1
deepseek	long-tail	3-of-3	4
deepseek	long-tail	3-of-3	7
eleutherai	long-tail	3-of-3	8
google-deepmind	pilot	2-of-3	2
hugging-face	long-tail	3-of-3	26
ibm	long-tail	3-of-3	10
kyutai	long-tail	3-of-3	4
lelapa-ai	long-tail	3-of-3	5
lg-ai-research	long-tail	3-of-3	7
liquid-ai	long-tail	3-of-3	1
maritaca-ai	long-tail	3-of-3	4
meta	pilot	2-of-3	27
microsoft	pilot	2-of-3	19
minimax	long-tail	3-of-3	7
mistral-ai	long-tail	3-of-3	10
moonshot-ai	long-tail	3-of-3	6
naver	long-tail	3-of-3	1
nous-research	long-tail	3-of-3	4
nvidia	long-tail	3-of-3	5
openai	pilot	2-of-3	25
openbmb	long-tail	3-of-3	8
reka-ai	long-tail	3-of-3	4
sakana-ai	long-tail	3-of-3	4
sarvam-ai	long-tail	3-of-3	4
snowflake	long-tail	3-of-3	4
stability-ai	long-tail	3-of-3	4
tencent	long-tail	3-of-3	7
tii	long-tail	3-of-3	5
upstage	long-tail	3-of-3	3
xai	long-tail	3-of-3	11
xiaomi	long-tail	3-of-3	4
zhipu-ai	long-tail	3-of-3	5

What those claims proposed to do
Kind	Count	Effect
Change	348	Advances a verification date in place against freshly fetched evidence; adds no record.
Add	4	Proposes a new record. None reached the dataset this run.

Not covered

Creator sites that publish no machine-readable model catalogue are swept only for the pages already in the approved source list; a new model announced solely in a blog post outside that list is not discovered.
Licence pages that require JavaScript to render could not be read as text, so osiApproved was left untouched wherever it was already recorded.
Serving-platform catalogues are deliberately out of scope: a platform offering a model states nothing about how its creator releases it.

Degraded discovery channels

openai-gpt-4-1-announcement — https://openai.com/index/gpt-4-1/: Fetch returned http-403, so no usable evidence was obtained. No claim rests on this channel; the records it backs keep their prior verification date rather than being advanced on stale evidence.
openai-gpt-5-6-launch — https://openai.com/index/gpt-5-6/: Fetch returned http-403, so no usable evidence was obtained. No claim rests on this channel; the records it backs keep their prior verification date rather than being advanced on stale evidence.
openai-gpt-5-6-price-update — https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/: Fetch returned http-403, so no usable evidence was obtained. No claim rests on this channel; the records it backs keep their prior verification date rather than being advanced on stale evidence.
openai-gpt-5-6-sol-preview — https://openai.com/index/previewing-gpt-5-6-sol/: Fetch returned http-403, so no usable evidence was obtained. No claim rests on this channel; the records it backs keep their prior verification date rather than being advanced on stale evidence.
openai-gpt-5-5-announcement — https://openai.com/index/introducing-gpt-5-5/: Fetch returned http-403, so no usable evidence was obtained. No claim rests on this channel; the records it backs keep their prior verification date rather than being advanced on stale evidence.
openai-gpt-5-4-announcement — https://openai.com/index/introducing-gpt-5-4/: Fetch returned http-403, so no usable evidence was obtained. No claim rests on this channel; the records it backs keep their prior verification date rather than being advanced on stale evidence.
alphabet-about — https://abc.xyz/: Fetch returned http-403, so no usable evidence was obtained. No claim rests on this channel; the records it backs keep their prior verification date rather than being advanced on stale evidence.

What was evaluated

Reviewers: 3
Verdicts cast: 1,056
Accepted by panel: 346
Rejected by panel: 6

Deterministic gates and required checks — 6 of 8 checks passed. Exit 0 is a pass; exit 2 means the gate could not run and is never treated as one.
Check	Scope	Exit	Result
`gate-evidence`	applied set, 44 bundles	0	PassEvery applied claim carries a url, a verbatim quote, a content hash and a fetch date. Two-sided: the same gate returned exit 1 over the unsplit bundles, naming exactly the six panel-rejected claims.
`gate-evidence`not required	unsplit bundles, control arm	1	FailThe failing arm that makes the passing arm readable. Refused three bundles for the six claims that missed their threshold.
`gate-source-approval`	applied set, 44 bundles	0	PassEvery cited origin is in the approved list at trunk 5f32f323d2.
`gate-source-approval`not required	fabricated unapproved origin, control arm	1	FailBuilt deliberately because the real DeepSeek claims rest on already-trusted origins and so exited 0, leaving the gate one-sided without it.
`gate-dataset`	web/src/data, 645 records	0	PassAll structural gates passed. It did not catch the model-fit staleness described in the caveats; only npm run validate did, so the two checks are not redundant.
`gate-scope`	branch against trunk merge-base	0	PassIn class, with four dataset documents changed. Exit 0 alone is ambiguous, so the changed count was read alongside it.
`npm run validate`	web, committed tree	0	Pass3189 tests across 139 files. On the uncommitted tree one assertion in scripts/licence-identity.test.ts fails because its instrument is git status --porcelain src/data, which cannot tell "the tool under test wrote" from "the tree was already dirty". Controlled both ways: clean tree 8 passed at exit 0, dirty tree that one assertion red.
`gate-ledger`	this entry	0	PassPrior entries untouched and in the same relative order; transcription reported false.

Applied over a recorded dissent

These met their threshold and were applied. The objection stands on the record and was not overruled.

apple-openelm-research-page--lastCheckedDateProvenanceThe apple-openelm-research-page quote is only footer/legal chrome ("Privacy Policy Terms of Use Legal Copyright 2026 Apple Inc.") and does not substantively evidence the OpenELM research page.
families--apple-openelm--verifiedAtProvenanceThe shown evidence for apple-openelm includes apple-openelm-research-page, whose quote is only footer/legal chrome ("Privacy Policy Terms of Use Legal Copyright 2026 Apple Inc."), so not every cited source is substantively evidenced.
organizations--apple--verifiedAtProvenanceThe shown evidence for the Apple organization includes apple-openelm-research-page, whose quote is only footer/legal chrome ("Privacy Policy Terms of Use Legal Copyright 2026 Apple Inc."), so not every cited source is substantively evidenced.
deepseek-v4-1-flash--addProvenanceThe model-card and Hub artefacts confirm the name, 552B backbone parameters, one-million-token context, image+text input with text output, activation figures, MIT licence name, repository files, and createdAt 2026-09-10. However the proposed release sets license.osiApproved=true, and none of the claim evidence or supplied artefacts is an Open Source Initiative source or states OSI approval; the provenance rubric requires OSI approval to be sourced from OSI, not inferred from the MIT licence name or license:mit metadata.
openai-models-catalog--lastCheckedDateProvenanceRejected by provenance and consistency and editorial. The openai-models-catalog quote only names GPT-Image-2.5 Sunburst and GPT-Image-2.5 Flare, so it does not substantively establish that the catalog still carries GPT-Image-2 as the subject claimed.
releases--openai-gpt-image-2--verifiedAtProvenanceRejected by provenance and consistency and editorial. The openai-models-catalog evidence shown for this verifiedAt update again quotes only GPT-Image-2.5 entries, so not every cited source is shown to still name the openai-gpt-image-2 subject.

Posted 340 edits

340 edits across 4 documents, a net change of 0 records.

Dataset documents this run changed
Document	Before	After	What changed
`web/src/data/families.json`	87	87	Every edit advances a verification date in place, so the record count is unchanged by construction.
`web/src/data/organizations.json`	44	44	Every edit advances a verification date in place, so the record count is unchanged by construction.
`web/src/data/releases.json`	129	129	Every edit advances a verification date in place, so the record count is unchanged by construction.
`web/src/data/sources.json`	296	296	Every edit advances a verification date in place, so the record count is unchanged by construction.

Each document links to the file as this run left it, not as it stands today.

Not posted 6 items

Accepted by the panel, then dropped

deepseek-v4-1-flash-model-card--addAccepted by the panel and deliberately not applied. It exists only to support deepseek-v4-1-flash--add, which the panel rejected 2/3 against a unanimous long-tail bar for asserting license.osiApproved with no OSI source. Applying the siblings alone would create a family with zero releases and two sources cited only by that empty family. The tranche also takes the home page to 1105251 critical raw bytes against a 1105000 ceiling, a 251-byte breach; widening web/asset-budgets.json is forbidden by ADR 0015 and outside the ADR 0003 class. Held, not rejected: the verdicts stand and the evidence is carried into the summary issue.Blocked by web/asset-budgets.jsondocs/adr/0015-asset-budget-measurements-are-in-class-ceilings-are-not.md
hugging-face-deepseek-v4-1-flash-hub-record--addAccepted by the panel and deliberately not applied. It exists only to support deepseek-v4-1-flash--add, which the panel rejected 2/3 against a unanimous long-tail bar for asserting license.osiApproved with no OSI source. Applying the siblings alone would create a family with zero releases and two sources cited only by that empty family. The tranche also takes the home page to 1105251 critical raw bytes against a 1105000 ceiling, a 251-byte breach; widening web/asset-budgets.json is forbidden by ADR 0015 and outside the ADR 0003 class. Held, not rejected: the verdicts stand and the evidence is carried into the summary issue.Blocked by web/asset-budgets.jsondocs/adr/0015-asset-budget-measurements-are-in-class-ceilings-are-not.md
deepseek-v4-1--addAccepted by the panel and deliberately not applied. It exists only to support deepseek-v4-1-flash--add, which the panel rejected 2/3 against a unanimous long-tail bar for asserting license.osiApproved with no OSI source. Applying the siblings alone would create a family with zero releases and two sources cited only by that empty family. The tranche also takes the home page to 1105251 critical raw bytes against a 1105000 ceiling, a 251-byte breach; widening web/asset-budgets.json is forbidden by ADR 0015 and outside the ADR 0003 class. Held, not rejected: the verdicts stand and the evidence is carried into the summary issue.Blocked by web/asset-budgets.jsondocs/adr/0015-asset-budget-measurements-are-in-class-ceilings-are-not.md
releases--openai-gpt-5--verifiedAtAccepted by the panel on sound freshly fetched evidence and withheld after applying revealed a whole-dataset invariant the claim could not see. web/src/data/validate.ts refuses a model-fit statement whose verifiedAt is older than a fact it rests on, and this release is pinned by such a statement still dated 2026-08-31. Advancing the release alone strands its guidance; advancing the guidance too would assert an editorial re-check no reviewer ever saw. Withholding errs toward understating verification, which is the recoverable direction. Trunk already held all four pinned releases back for the same reason, so this is a standing constraint the scouting model does not yet represent.Blocked by web/src/data/validate.tsweb/src/data/model-fit-statements.json
releases--anthropic-claude-mythos-5--verifiedAtAccepted by the panel on sound freshly fetched evidence and withheld after applying revealed a whole-dataset invariant the claim could not see. web/src/data/validate.ts refuses a model-fit statement whose verifiedAt is older than a fact it rests on, and this release is pinned by such a statement still dated 2026-08-31. Advancing the release alone strands its guidance; advancing the guidance too would assert an editorial re-check no reviewer ever saw. Withholding errs toward understating verification, which is the recoverable direction. Trunk already held all four pinned releases back for the same reason, so this is a standing constraint the scouting model does not yet represent.Blocked by web/src/data/validate.tsweb/src/data/model-fit-statements.json
releases--anthropic-claude-haiku-4-5--verifiedAtAccepted by the panel on sound freshly fetched evidence and withheld after applying revealed a whole-dataset invariant the claim could not see. web/src/data/validate.ts refuses a model-fit statement whose verifiedAt is older than a fact it rests on, and this release is pinned by such a statement still dated 2026-08-31. Advancing the release alone strands its guidance; advancing the guidance too would assert an editorial re-check no reviewer ever saw. Withholding errs toward understating verification, which is the recoverable direction. Trunk already held all four pinned releases back for the same reason, so this is a standing constraint the scouting model does not yet represent.Blocked by web/src/data/validate.tsweb/src/data/model-fit-statements.json

What this run does not prove

7 source channels came back degraded and carry no claim this run; the records they back keep their prior verification date rather than being advanced on stale evidence.
Around 20 licence pages could not be read as text, so no osiApproved field was touched. A licence name is not OSI status, and inferring one from the other is a rejection under the provenance rubric.
Scout defect, marker prefix matching: verify.mjs compared alphanumeric-squashed strings, so a marker that is a strict prefix of a longer version matched it — GPT-Image-2 inside GPT-Image-2.5 — a false-present verification, which is the dangerous direction. Measured scope across 249 verified rows with a three-way classifier control: exactly 1 genuine defect, and the panel had already caught it with all three rubrics independently. Six other prefix hits are benign family-inside-member matches.
Scout defect, evidence truncation: make-review-payloads.mjs truncated evidence to two items while evidenceCount reported the true total. The provenance prompt disclosed the truncation and the consistency prompt did not, so one reviewer correctly rejected on the mismatch. The payload was regenerated untruncated and re-judged — legitimate because the input was defective, not because the answer was inconvenient.
The three-rubric panel is three prompts, not three minds: instances of one model family reading one page share a failure mode. The majority buys independence of reasoning, not of training, and a source that is itself wrong can carry all three.
gate-dataset passed at exit 0 over the tree that npm run validate then rejected. The deterministic gates do not subsume the test suite, and a run that stopped at the gates would have published a dataset that fails its own validator.
The asset-budget headroom on the home page is 2,225 bytes against a 1,105,000 ceiling. This run is byte-neutral, but the next run that adds any record will meet the same ceiling the DeepSeek tranche met.

Follow-ups — proposed, not fixed

Teach the scouting model the model-fit pin: a release whose verifiedAt is referenced by a model-fit statement cannot be advanced alone. Today the coupling is discovered only by applying and running the validator, which is late and wastes a panel cycle.
Fix the marker comparison in the scout to require a token boundary rather than a substring, so a prefix cannot verify a longer version. The panel caught this instance; the next one may be in a bundle with a thinner evidence trail.
scripts/licence-identity.test.ts asserts that the sweep writes no dataset file using git status --porcelain src/data, which also fails whenever the tree is merely dirty. Comparing file hashes around the call would measure what the test means.
modeltree-publish/SKILL.md says to prepend the ledger entry, but entries 14 onward were appended and runsNewestFirst() sorts rather than trusting array order. The wording is stale; the behaviour is fine.

2026-09-21-2f9a3a

Verification dates refreshed across 45 creator bundles; four accepted releases held at the asset budget

Scope requested: Every creator in the catalogue. No creator was skipped, so found.unswept is empty; the eight bundles whose discovery channels answered 403 were still scouted through their other catalogued sources.PublishedSep 21, 2026526 edits posted · 4 items withheld

A full-catalogue pass: 45 bundles scouted from primary sources, 325 pages fetched and hashed, 530 claims put to a three-rubric panel, and 526 applied. Every applied claim advances a verification date in place against freshly fetched evidence, so the dataset gained no record and lost none. Four panel-accepted claims adding the two GPT-Image-2.5 releases were held back rather than applied, because landing them breaches the home page asset budget and repairing that would mean editing a file outside the class this run is allowed to touch.

PreflightRanTrunk anchor resolved to 32fae15cba0b9884d18935353f84fc9b681627b5, reflog-dated 2026-09-19 08:43:32 -0400, and re-read against the live remote at report time. Six open pull requests, none touching web/src/data.
ScoutRan325 pages fetched, 316 returning usable bytes; each stored with a sha256 taken at fetch time and every quote sliced back out of those bytes. A coverage gap was found by audit rather than by a gate: eight publishers carry no organizationId, so 25 fetched sources initially belonged to no bundle. A 45th shared long-tail bundle now covers them.
ReviewRan1590 verdicts from three rubrics, each blind to the others and to the scout's reasoning. Certified against a seeded 8-claim control before the real tally was read, because a sound panel and a rubber-stamping one produce identical all-accept tallies.
GatesRanNine checks, every exit code read unpiped, each with a control that came back the other way. Exit 2 was treated as a failure throughout and no gate was skipped, overridden or edited to pass.
PublishRan526 claims applied across four documents as 526 insertions and 526 deletions, then auto-merged on a green web-ci. Four accepted claims were held back and are recorded under withheld.
DeployRanThe GitHub Pages deployment for the merge commit was confirmed against the full 40-character SHA, with a differing control SHA establishing that the query discriminates.

What was found

Scouts: 45
Pages fetched and hashed: 325
Claims proposed: 530

Claims per creator bundle, with the review threshold its profile set
Creator	Policy	Threshold	Claims
01-ai	long-tail	3-of-3	8
ai-singapore	long-tail	3-of-3	5
ai2	long-tail	3-of-3	9
ai21-labs	long-tail	3-of-3	9
aleph-alpha	long-tail	3-of-3	5
alibaba-cloud	pilot	2-of-3	18
amazon	pilot	2-of-3	14
anthropic	pilot	2-of-3	33
apple	long-tail	3-of-3	11
baidu	long-tail	3-of-3	5
bytedance-seed	long-tail	3-of-3	7
cohere	long-tail	3-of-3	11
databricks	long-tail	3-of-3	5
deepseek	long-tail	3-of-3	10
eleutherai	long-tail	3-of-3	10
google-deepmind	pilot	2-of-3	14
hugging-face	long-tail	3-of-3	32
ibm	long-tail	3-of-3	17
kyutai	long-tail	3-of-3	5
lelapa-ai	long-tail	3-of-3	6
lg-ai-research	long-tail	3-of-3	9
liquid-ai	long-tail	3-of-3	5
maritaca-ai	long-tail	3-of-3	6
meta	pilot	2-of-3	39
microsoft	pilot	2-of-3	24
minimax	long-tail	3-of-3	12
mistral-ai	long-tail	3-of-3	19
moonshot-ai	long-tail	3-of-3	8
naver	long-tail	3-of-3	6
nous-research	long-tail	3-of-3	5
nvidia	long-tail	3-of-3	13
openai	pilot	2-of-3	37
openbmb	long-tail	3-of-3	13
reka-ai	long-tail	3-of-3	5
sakana-ai	long-tail	3-of-3	6
sarvam-ai	long-tail	3-of-3	5
shared	long-tail	3-of-3	25
snowflake	long-tail	3-of-3	5
stability-ai	long-tail	3-of-3	10
tencent	long-tail	3-of-3	9
tii	long-tail	3-of-3	9
upstage	long-tail	3-of-3	4
xai	long-tail	3-of-3	11
xiaomi	long-tail	3-of-3	5
zhipu-ai	long-tail	3-of-3	6

What those claims proposed to do
Kind	Count	Effect
Change	526	Verification dates advanced in place against freshly fetched, hashed evidence. Byte-neutral: each swaps one ten-character date for another.
Add	4	New records proposed. All four were held back at the budget gate and none reached the dataset.

Not covered

Ten of 325 fetches yielded nothing usable: nine returned HTTP 403 (seven catalogued sources plus the OpenAI and xAI announcement feeds), and zhipu-ai-glm-4-5-announcement returned HTTP 200 with zero extractable characters. Those sources kept their existing verification dates rather than being advanced on a page nobody read. Note that the 200-with-no-text case is recorded as a successful fetch by http status alone, so the usable count is 315 rather than the 316 that status suggests.
Anthropic verification dates on four records were deliberately not advanced, because advancing them would strand model-fit statements whose own verifiedAt is older, which validate.ts refuses. Recorded here rather than worked around.
Two Gemini documentation pages were served localized by content negotiation (Hindi and Portuguese). The quotes are verbatim and still identifying, but the channel was degraded and is reported as such.

Degraded discovery channels

openai — openai-gpt-4-1-announcement: https://openai.com/index/gpt-4-1/ returned HTTP 403. No user agent was spoofed to work around it, so this creator's catalogued source went unread this pass.
openai — openai-gpt-5-6-launch: https://openai.com/index/gpt-5-6/ returned HTTP 403. No user agent was spoofed to work around it, so this creator's catalogued source went unread this pass.
openai — openai-gpt-5-6-price-update: https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/ returned HTTP 403. No user agent was spoofed to work around it, so this creator's catalogued source went unread this pass.
openai — openai-gpt-5-6-sol-preview: https://openai.com/index/previewing-gpt-5-6-sol/ returned HTTP 403. No user agent was spoofed to work around it, so this creator's catalogued source went unread this pass.
openai — openai-gpt-5-5-announcement: https://openai.com/index/introducing-gpt-5-5/ returned HTTP 403. No user agent was spoofed to work around it, so this creator's catalogued source went unread this pass.
openai — openai-gpt-5-4-announcement: https://openai.com/index/introducing-gpt-5-4/ returned HTTP 403. No user agent was spoofed to work around it, so this creator's catalogued source went unread this pass.
shared — alphabet-about: https://abc.xyz/ returned HTTP 403. No user agent was spoofed to work around it, so this creator's catalogued source went unread this pass.
zhipu-ai — zhipu-ai-glm-4-5-announcement: https://z.ai/blog/glm-4.5 answered HTTP 200 with 598 bytes of markup and zero extractable characters. The fetch is recorded as ok, but a page that answers with nothing readable is not a page that was read, so nothing was verified against it.
openai — https://openai.com/news/ (discovery feed): https://openai.com/news/ returned HTTP 403. No user agent was spoofed to work around it, so this creator's announcement feed went unread this pass.
xai — https://x.ai/news (discovery feed): https://x.ai/news returned HTTP 403. No user agent was spoofed to work around it, so this creator's announcement feed went unread this pass.

What was evaluated

Reviewers: 3
Verdicts cast: 1,590
Accepted by panel: 530
Rejected by panel: 0

Deterministic gates and required checks — 9 of 9 checks passed. Exit 0 is a pass; exit 2 means the gate could not run and is never treated as one.
Check	Scope	Result
`gate-evidence`	45 bundles, 530 claims	PassEvery claim carries a fetched page, a content hash taken at fetch time, and a quote re-sliced from those stored bytes. Controlled in the same session: an all-reject claim and a claim with no verdicts each exited 1.
`gate-source-approval`	45 bundles	PassAll 45 agree on a single anchor matching trunk. Exactly two proposed sources run-wide, both held back, so this run approves no new source and no new publisher. An untrusted origin exited 1.
`gate-dataset`	645 records across 4 documents	PassSchema, referential integrity and freshness coherence over the applied tree. Zero failures.
`gate-scope`	committed tip against the computed merge-base	PassFour changed paths, all inside web/src/data. A stray out-of-class file produced passed:false in the same session, so the pass is a reading rather than a default.
`gate-reversals`	web/src/data	PassNo field was advanced and then walked back. Baselined at exit 0 on the clean anchor tree before the change, so a non-zero afterwards would have been attributable to this run.
`gate-ledger`	branch against the anchor	PassRecord counts per declared document match the diff. Ran before this entry existed and correctly failed, which is what makes the later pass readable.
`npm run validate`	web/	Pass139 test files, 3,189 tests, all passing on the committed tree.
`ci-preflight`	repository root	PassSelects the checks this diff actually triggers: web-ci, skills-ci and source-link-health-tests all passed locally.
`panel control`	8 seeded claims through the live panel	PassFour sound claims accepted 3-of-3; four planted defects rejected at both thresholds. The arms separated, so the all-accept tally above measures something rather than nothing.

Applied over a recorded dissent

These met their threshold and were applied. The objection stands on the record and was not overruled.

releases-openai-gpt-image-2-5-sunburst-add-2026-09-21Editorial and entity boundariesREJECT (add release): the summary imports OpenAI's verbatim first-person marketing superlative 'Our most capable model for image generation and editing' as the dataset's own unattributed characterization. This asserts a rank/winner in the dataset's phrasing (rubric: no composite score, rank, or winner in the data or phrasing) and breaks the dataset norm of converting vendor superlatives to third-person attribution or a marked quotation — the sibling live record openai-gpt-image-2 ('the current image model') and the same-run Flare add (neutral tagline) both do this correctly, making Sunburst the outlier. Fix: reword to a neutral/attributed form (e.g. drop the superlative like gpt-image-2, or 'OpenAI's catalogue lists it under Image as its most capable image model'). Naming/entity/modalities are otherwise fine.

Posted 526 edits

526 edits across 4 documents, a net change of 0 records.

Dataset documents this run changed
Document	Before	After	What changed
`web/src/data/sources.json`	296	296	288 field value(s) advanced in place. Record count unchanged, and measured against the committed parent at 32fae15cba rather than assumed.
`web/src/data/organizations.json`	44	44	42 field value(s) advanced in place. Record count unchanged, and measured against the committed parent at 32fae15cba rather than assumed.
`web/src/data/families.json`	87	87	81 field value(s) advanced in place. Record count unchanged, and measured against the committed parent at 32fae15cba rather than assumed.
`web/src/data/releases.json`	129	129	115 field value(s) advanced in place. Record count unchanged, and measured against the committed parent at 32fae15cba rather than assumed.

Each document links to the file as this run left it, not as it stands today.

Not posted 4 items

Accepted by the panel, then dropped

sources-openai-gpt-image-2-5-sunburst-docs-add-2026-09-21Accepted by the panel at 3/3 (needed 2) and deliberately not applied. Applying the four GPT-Image-2.5 claims together takes the home page to 1,107,990 critical raw bytes against a 1,105,000 ceiling. Landing them needs web/asset-budgets.json widened, which is outside the ADR 0003 qualifying class and which that test forbids in its own text, so the budget decision is left to a human. Held, not rejected: the verdicts stand and the evidence is carried into a follow-up issue.Blocked by web/tests/build/asset-budgets.test.tsweb/asset-budgets.json
sources-openai-gpt-image-2-5-flare-docs-add-2026-09-21Accepted by the panel at 3/3 (needed 2) and deliberately not applied. Applying the four GPT-Image-2.5 claims together takes the home page to 1,107,990 critical raw bytes against a 1,105,000 ceiling. Landing them needs web/asset-budgets.json widened, which is outside the ADR 0003 qualifying class and which that test forbids in its own text, so the budget decision is left to a human. Held, not rejected: the verdicts stand and the evidence is carried into a follow-up issue.Blocked by web/tests/build/asset-budgets.test.tsweb/asset-budgets.json
releases-openai-gpt-image-2-5-sunburst-add-2026-09-21Accepted by the panel at 2/3 (needed 2) and deliberately not applied. Applying the four GPT-Image-2.5 claims together takes the home page to 1,107,990 critical raw bytes against a 1,105,000 ceiling. Landing them needs web/asset-budgets.json widened, which is outside the ADR 0003 qualifying class and which that test forbids in its own text, so the budget decision is left to a human. Held, not rejected: the verdicts stand and the evidence is carried into a follow-up issue.Blocked by web/tests/build/asset-budgets.test.tsweb/asset-budgets.json
releases-openai-gpt-image-2-5-flare-add-2026-09-21Accepted by the panel at 3/3 (needed 2) and deliberately not applied. Applying the four GPT-Image-2.5 claims together takes the home page to 1,107,990 critical raw bytes against a 1,105,000 ceiling. Landing them needs web/asset-budgets.json widened, which is outside the ADR 0003 qualifying class and which that test forbids in its own text, so the budget decision is left to a human. Held, not rejected: the verdicts stand and the evidence is carried into a follow-up issue.Blocked by web/tests/build/asset-budgets.test.tsweb/asset-budgets.json

What this run does not prove

A green run proves the applied dates match pages fetched today. It does not prove those pages are themselves correct, nor that a creator published something this pass missed.
All 530 claims were accepted. That tally is only readable because the seeded control separated its arms first: four planted defects were rejected at both thresholds while four sound claims passed 3-of-3.
One claim carried a recorded editorial objection and was applied anyway on a 2-of-3 pilot threshold. Raising a threshold after seeing the vote is as much a bypass as lowering it, so the dissent is published rather than acted on. It is one of the four claims held back for the unrelated budget reason.
The 25-source coverage gap was caught by an audit of my own bundling, not by any gate. No gate here checks that every fetched source reached a bundle, so a similar gap would pass unnoticed next run.
My first audit tool reported every source as unfetched because it read an id field that is named sourceId. That was a broken instrument rather than a finding, and it is recorded so the next reader does not trust a similar sweep without a control.
The home page sits at 1,102,773 raw bytes against a 1,105,000 ceiling: 2,227 bytes of headroom, 99.8% of budget. Any dataset growth at all now reddens the build, which is why four accepted claims could not land.
web/scripts/licence-identity.test.ts asserts that git status --porcelain src/data is empty, so it reddens on any uncommitted dataset edit and cannot distinguish a licence tool writing to src/data from a tree that already held refresh edits. It passes here and in CI; the ambiguity is a latent defect, not a failure this run caused.

Follow-ups — proposed, not fixed

The four held GPT-Image-2.5 claims keep their panel verdicts and their fetched, hashed evidence. They need a human decision on the home page asset budget before they can land, and the Sunburst summary should be reworded to drop the imported first-person superlative before it does.
The home page budget has 2,227 bytes of headroom. Either the critical payload shrinks or the ceiling is revisited deliberately; the next refresh that adds a record will hit this again.
No gate asserts that every successfully fetched source reached a bundle. This run needed a hand audit to notice 25 that had not.

2026-09-19-4f415a

Full sweep of all 44 creators, 523 edits published, and a panel whose whole discriminating power came from one rubric

Scope requested: All 44 creators in organizations.json, with no creator narrowed out, plus one shared bundle for cross-creator sources that belong to no single organisation - 45 bundles in total. found.unswept is empty because every catalogued creator was reached, not because nothing was recorded there: the set of bundled creators was reconciled against organizations.json and the difference is zero. 335 page fetches were attempted and 325 returned HTTP 200; the rest are recorded as degraded channels against the creator whose channel failed, because a creator that was looked at through a channel that went dark is in a different state from a creator nobody reached.PublishedSep 19, 2026523 edits posted · 10 items withheld

Every catalogued creator was scouted from primary sources in one pass. 335 pages were fetched and hashed in this run; 528 claims were proposed, a three-rubric panel cast 1584 verdicts, and all 528 met their per-profile threshold. 523 reached the dataset. Every claim is a re-verification carrying a verification date forward on a fact re-read at its source, so no record was added or removed and every record count is unchanged. Five accepted claims were then refused by npm run validate, which found they would have left a model-fit statement verified ahead of the release or family fields it rests on; they are published as withheld under dropped-after-acceptance rather than discarded. All five had been rejected by the consistency rubric and carried by the other two, and provenance and editorial cast byte-identical all-accept vectors across every claim - so the lone dissenting rubric was the only thing in the panel that discriminated, and on all five it was right.

PreflightRanClean tree at acae4097b4, which was also refs/remotes/origin/main and was confirmed current against the remote at the moment of use; its reflog dates the write to 2026-09-18 12:03:35 -0400. No refresh pull request was left open from a previous run - six pull requests were open and each was checked for dataset overlap, with zero files under web/src/data between them. Both invocation forms were probed for every tool rather than either being assumed: bare npm, npx and drydock are all refused by this host's PowerShell execution policy, while the .cmd forms run. That is installed and blocked, not absent, and the execution policy was not changed. npm ci was used for install and the lockfile was verified unmodified afterwards.
ScoutRan335 pages fetched across 45 bundles, each stored with its sha256 content hash and fetch date. Every quote was sliced programmatically out of the fetched bytes and re-checked against those bytes before the claim was built; no quote was recalled and no search snippet was treated as evidence. 528 claims proposed. 29 facts were held rather than claimed because their evidence was not good enough. Three separate blind-matcher defects were found and fixed in this run's own tooling before any claim was trusted, each caught by a control that landed on the failing path rather than by a check that merely returned nothing.
ReviewRanThree independent rubrics per claim, each blind to the other two and to the scout's reasoning, which was stripped from the review input. 1584 verdicts over 528 claims, 2-of-3 for the 7 pilot bundles and unanimous 3-of-3 for the 38 long-tail bundles. By rubric: provenance 528 accept / 0 reject, consistency 523 / 5, editorial 528 / 0. A validator flagged the verdict join as broken; investigating rather than accepting it found the validator itself was blind - it read c.id where the field is claimId - and fixing it surfaced a real defect in the scout, 6 colliding claim ids. Verdicts were then joined to claims by index, proved exact against a deliberate-swap control that correctly reported 2 misalignments, so no vote moved. No verdict was reversed and no threshold was lowered.
GatesRanSix required gates, every exit code read from an unpiped invocation, and each run with a two-arm control in the same session so that a pass is readable rather than merely green. All six passed. gate-evidence found a real defect in the run that was invoking it. npm run validate refused five claims the panel had accepted, and the gate was right: the dissenting rubric had objected to all five. No gate was skipped, no threshold lowered, no override used and nothing was edited to make a gate pass.
PublishRanDataset committed before this entry was written, so the entry could name its own pull request rather than guess a number - the ordering every recent published run has followed, because guessing would put a falsehood in the permanent record and force-pushing to hide it would be a bypass. The pull request carries the full evidence trail and auto-merges on a green web-ci; it was not merged by hand, no branch protection was bypassed and nothing was pushed to main directly.
DeployRanPages deploy confirmed after the merge against the full 40-character merge commit SHA, with a control SHA selected to differ from it so the two arms had to disagree before either was read.

What was found

Scouts: 44
Pages fetched and hashed: 335
Claims proposed: 528

Claims per creator bundle, with the review threshold its profile set
Creator	Policy	Threshold	Claims
01-ai	long-tail	3-of-3	8
ai-singapore	long-tail	3-of-3	5
ai2	long-tail	3-of-3	9
ai21-labs	long-tail	3-of-3	7
aleph-alpha	long-tail	3-of-3	5
alibaba-cloud	pilot	2-of-3	18
amazon	pilot	2-of-3	14
anthropic	pilot	2-of-3	36
apple	long-tail	3-of-3	11
baidu	long-tail	3-of-3	5
bytedance-seed	long-tail	3-of-3	5
cohere	long-tail	3-of-3	11
databricks	long-tail	3-of-3	5
deepseek	long-tail	3-of-3	10
eleutherai	long-tail	3-of-3	10
google-deepmind	pilot	2-of-3	13
hugging-face	long-tail	3-of-3	32
ibm	long-tail	3-of-3	17
kyutai	long-tail	3-of-3	5
lelapa-ai	long-tail	3-of-3	6
lg-ai-research	long-tail	3-of-3	9
liquid-ai	long-tail	3-of-3	5
maritaca-ai	long-tail	3-of-3	6
meta	pilot	2-of-3	40
microsoft	pilot	2-of-3	24
minimax	long-tail	3-of-3	12
mistral-ai	long-tail	3-of-3	19
moonshot-ai	long-tail	3-of-3	8
naver	long-tail	3-of-3	6
nous-research	long-tail	3-of-3	5
nvidia	long-tail	3-of-3	13
openai	pilot	2-of-3	35
openbmb	long-tail	3-of-3	13
reka-ai	long-tail	3-of-3	5
sakana-ai	long-tail	3-of-3	6
sarvam-ai	long-tail	3-of-3	5
shared	long-tail	3-of-3	25
snowflake	long-tail	3-of-3	5
stability-ai	long-tail	3-of-3	10
tencent	long-tail	3-of-3	9
tii	long-tail	3-of-3	9
upstage	long-tail	3-of-3	4
xai	long-tail	3-of-3	12
xiaomi	long-tail	3-of-3	5
zhipu-ai	long-tail	3-of-3	6

What those claims proposed to do
Kind	Count	Effect
Change	528	Field values advanced in place on records already in the dataset. No record added or removed.

Not covered

Benchmark results and usage observations were not scouted this run. The sweep covered creators, families, releases and their sources only.
3 source(s) returned HTTP 200 with too little text to quote from, so they yielded no claims even though the fetch itself succeeded. They are listed as degraded channels, because a page that answers with nothing readable is not a page that was read.
check-bundle-pairing passed, and the pass measured nothing: it checks that a newly added source is paired with a record citing it, and this run proposed zero source-add claims, so it was structurally silent. It is recorded as a vacuous pass rather than as evidence.

Degraded discovery channels

long-tail — xai-news: 1 fetch(es) to x.ai returned HTTP 403. Reported as a channel this run could not hear from; no user agent was spoofed and the block was not worked around.
openai — openai-gpt-4-1-announcement: 7 fetch(es) to openai.com returned HTTP 403. Reported as a channel this run could not hear from; no user agent was spoofed and the block was not worked around.
shared — alphabet-about: 1 fetch(es) to abc.xyz returned HTTP 403. Reported as a channel this run could not hear from; no user agent was spoofed and the block was not worked around.
bytedance-seed — bytedance-seed-site: HTTP 200 with 14 characters of extractable text. Treated as incomplete, not as a page that said nothing.
meta — meta-ai-site: HTTP 200 with 49 characters of extractable text. Treated as incomplete, not as a page that said nothing.
zhipu-ai — zhipu-ai-glm-4-5-announcement: HTTP 200 with 0 characters of extractable text. Treated as incomplete, not as a page that said nothing.

What was evaluated

Reviewers: 3
Verdicts cast: 1,584
Accepted by panel: 528
Rejected by panel: 0

Deterministic gates and required checks — 6 of 6 checks passed. Exit 0 is a pass; exit 2 means the gate could not run and is never treated as one.
Check	Scope	Result
`gate-evidence`	45 claim bundles, 528 claims	PassPassed on all 45 bundles. It first refused 3 bundles, correctly: this run's own scout had minted 6 duplicate claim ids because its id template omitted the collection, and a family and a release may share an id. The ids were requalified with their collection and nothing was relaxed. Controlled in the same session: flipping one verdict to reject on a long-tail claim exits 1 ("only reached 2 of 3 required accepts"), and emptying a rationale exits 1 ("gave no rationale, so its verdict is unauditable"). The gate derives the threshold from the reviewed-profile set rather than reading the bundle, and it agreed with this run: 7 pilot creators, 38 long-tail.
`gate-source-approval`	45 claim bundles	PassPassed on all 45 bundles. It anchors on git merge-base HEAD refs/remotes/origin/main and reads the dataset from git rather than the working tree, so running it could not let this run vouch for its own writes. Controlled: repointing a citation at a fabricated origin exits 1. No new publisher was approved by this run.
`gate-dataset`	645 records in web/src/data	PassPassed over 645 records on the applied tree. Controlled: repointing a release familyId at a non-existent family exits 1 ("does not resolve to a family"). Note its boundary, measured rather than assumed - it does not check field formats, so a malformed date passes it at exit 0 and is caught by the Zod schema under npm run validate instead. The two are complementary, and neither subsumes the other.
`gate-scope`	the committed tip against the merge base acae4097b4	PassIn class: exactly 4 dataset documents changed - families.json, organizations.json, releases.json, sources.json. Nothing outside web/src/data was touched. Controlled: appending one line to README.md exits 1 with "OUT OF CLASS". Run both before and after the dataset commit, because HEAD is what auto-merges.
`gate-reversals`	web/src/data	Pass9 of 18 rejected records in the dataset are annotated by rejection-reversals.json. The gate reports its own coverage limit honestly: 264 of 282 rejections name no record in their detail, so it cannot tell whether those were reversed, and it says not checked is not passed rather than reporting them as clean. rejection-reversals.json was not touched by this run.
`npm run validate`	the committed tree, from web/	Pass139 test files, 3,189 tests, 0 failures; astro check reports 0 errors, 0 warnings, 0 hints across 287 files. Exit code read from an unpiped invocation: a first reading taken through a pipeline reported exit 1 while every test passed, which is the pipeline corrupting the status rather than a failing run, and the unpiped re-read is the one recorded here. This gate is also what refused the five claims the panel had accepted; they are published as withheld rather than dropped in silence.

Applied over a recorded dissent

These met their threshold and were applied. The objection stands on the record and was not overruled.

anthropic-claude-4-5-reverify-2026-09-19Cross-source consistencyAdvancing families anthropic-claude-4-5 verifiedAt from 2026-08-31 to 2026-09-19 strands dependent model-fit statement(s) fit-claude-haiku-4-5-family-legacy (verifiedAt 2026-08-31): validate.ts line 433 requires statement.verifiedAt >= the verifiedAt of every fact it cites, so the statement's 2026-08-31 would fall before its cited fact's 2026-09-19 ("was verified before"). No model-fit statement claim in this bundle re-verifies it, so the advance breaks the build gate.
anthropic-claude-haiku-4-5-reverify-2026-09-19Cross-source consistencyAdvancing releases anthropic-claude-haiku-4-5 verifiedAt from 2026-08-31 to 2026-09-19 strands dependent model-fit statement(s) fit-claude-haiku-4-5-release-current (verifiedAt 2026-08-31): validate.ts line 433 requires statement.verifiedAt >= the verifiedAt of every fact it cites, so the statement's 2026-08-31 would fall before its cited fact's 2026-09-19 ("was verified before"). No model-fit statement claim in this bundle re-verifies it, so the advance breaks the build gate.
anthropic-claude-mythos-5-reverify-2026-09-19Cross-source consistencyAdvancing releases anthropic-claude-mythos-5 verifiedAt from 2026-08-31 to 2026-09-19 strands dependent model-fit statement(s) fit-claude-mythos-5-limited-availability (verifiedAt 2026-08-31): validate.ts line 433 requires statement.verifiedAt >= the verifiedAt of every fact it cites, so the statement's 2026-08-31 would fall before its cited fact's 2026-09-19 ("was verified before"). No model-fit statement claim in this bundle re-verifies it, so the advance breaks the build gate.
meta-llama-4-scout-reverify-2026-09-19Cross-source consistencyAdvancing releases meta-llama-4-scout verifiedAt from 2026-08-31 to 2026-09-19 strands dependent model-fit statement(s) fit-llama-4-scout-self-hosting (verifiedAt 2026-08-31), fit-llama-4-scout-mau-threshold (verifiedAt 2026-08-31), fit-llama-4-scout-context-window (verifiedAt 2026-08-31): validate.ts line 433 requires statement.verifiedAt >= the verifiedAt of every fact it cites, so the statement's 2026-08-31 would fall before its cited fact's 2026-09-19 ("was verified before"). No model-fit statement claim in this bundle re-verifies it, so the advance breaks the build gate.
openai-gpt-5-reverify-2026-09-19Cross-source consistencyAdvancing releases openai-gpt-5 verifiedAt from 2026-08-31 to 2026-09-19 strands dependent model-fit statement(s) fit-gpt-5-superseded (verifiedAt 2026-08-31): validate.ts line 433 requires statement.verifiedAt >= the verifiedAt of every fact it cites, so the statement's 2026-08-31 would fall before its cited fact's 2026-09-19 ("was verified before"). No model-fit statement claim in this bundle re-verifies it, so the advance breaks the build gate.

Posted 523 edits

523 edits across 4 documents, a net change of 0 records.

Dataset documents this run changed
Document	Before	After	What changed
`web/src/data/sources.json`	296	296	287 field value(s) advanced in place. Record count unchanged, and measured against the committed parent rather than assumed.
`web/src/data/organizations.json`	44	44	41 field value(s) advanced in place. Record count unchanged, and measured against the committed parent rather than assumed.
`web/src/data/families.json`	87	87	81 field value(s) advanced in place. Record count unchanged, and measured against the committed parent rather than assumed.
`web/src/data/releases.json`	129	129	115 field value(s) advanced in place. Record count unchanged, and measured against the committed parent rather than assumed.

Each document links to the file as this run left it, not as it stands today.

Not posted 10 items

Accepted by the panel, then dropped

meta-llama-4-scout-reverify-2026-09-19Panel accepted 2-of-3 (pilot) over the consistency rubric's reject. validateDataset then refused it: model fit statements fit-llama-4-scout-self-hosting, -mau-threshold and -context-window would have been verified before the release fields accessType, license and contextWindow they rest on.Blocked by npm run validateweb/src/data/model-fit-statements.json
anthropic-claude-mythos-5-reverify-2026-09-19Panel accepted 2-of-3 (pilot) over the consistency rubric's reject. validateDataset refused it: fit-claude-mythos-5-limited-availability would have been verified before the release fields status and intendedUse it rests on.Blocked by npm run validateweb/src/data/model-fit-statements.json
anthropic-claude-haiku-4-5-reverify-2026-09-19Panel accepted 2-of-3 (pilot) over the consistency rubric's reject. validateDataset refused it: fit-claude-haiku-4-5-release-current would have been verified before the release field status it rests on.Blocked by npm run validateweb/src/data/model-fit-statements.json
anthropic-claude-4-5-reverify-2026-09-19Panel accepted 2-of-3 (pilot) over the consistency rubric's reject. validateDataset refused it: fit-claude-haiku-4-5-family-legacy would have been verified before the family field status it rests on. The release twin sharing this claimId is a separate record and was decided separately.Blocked by npm run validateweb/src/data/model-fit-statements.json
openai-gpt-5-release-reverify-2026-09-19Panel accepted 2-of-3 (pilot) over the consistency rubric's reject. validateDataset refused it: fit-gpt-5-superseded would have been verified before the release field status it rests on. The families twin sharing this claimId carries no dependent fit fact, was accepted by all three rubrics, and is applied.Blocked by npm run validateweb/src/data/model-fit-statements.json

Verification date deliberately held back

verification-hold-source10 fact(s) held: page http-error 403 (and 9 more of the same kind). Unknown stays unknown; a guessed field is worse than an empty one.Blocked by openai-gpt-4-1-announcementopenai-gpt-5-6-launchopenai-gpt-5-6-price-updateopenai-gpt-5-6-sol-previewopenai-gpt-5-5-announcementopenai-gpt-5-4-announcementgoogle-gemini-2-5-reportalphabet-about
verification-hold-releases11 fact(s) held: cites a source that did not fetch cleanly (openai-gpt-4-1-announcement:http-error) (and 10 more of the same kind). Unknown stays unknown; a guessed field is worse than an empty one.Blocked by openai-gpt-4-1-2025-04-14openai-gpt-4-1-mini-2025-04-14openai-gpt-4-1-nano-2025-04-14openai-gpt-5-6-solopenai-gpt-5-6-terraopenai-gpt-5-6-lunaopenai-gpt-5-5openai-gpt-5-4
verification-hold-families5 fact(s) held: cites a source that did not fetch cleanly (openai-gpt-4-1-announcement:http-error) (and 4 more of the same kind). Unknown stays unknown; a guessed field is worse than an empty one.Blocked by openai-gpt-4-1openai-gpt-5-6openai-gpt-5-5openai-gpt-5-4zhipu-ai-glm-4-5
verification-hold-organizations3 fact(s) held: cites a source that did not fetch cleanly (openai-gpt-4-1-announcement:http-error, openai-gpt-5-6-launch:http-error) (and 2 more of the same kind). Unknown stays unknown; a guessed field is worse than an empty one.Blocked by openaizhipu-aibytedance-seed

Blocked by policy before it could run

xai-grok-voice-transcribe-2-0-tranchexAI announced Grok Voice Transcribe 2.0 on 2026-09-17 and the release was verified against fetched, hashed bytes. Adding the record measured 2,596 bytes against 2,225 bytes of remaining home-route critical-path headroom, which is 371 bytes over the ceiling. A ceiling is out of the auto-merging class under ADR 0015, so the run withholds the record rather than raising the limit. This is the second consecutive tranche blocked by this one ceiling.Blocked by web/asset-budgets.json criticalMaxRaw for the home route

What this run does not prove

Two of the three rubrics returned byte-identical verdict vectors. Provenance and editorial each accepted all 528 claims, so the entire discriminating power of this panel came from the consistency rubric alone. A three-reviewer panel that behaves like a one-reviewer panel is a weaker check than its name suggests, and this run is the clearest evidence of that yet recorded here.
On all five claims where the panel split, the lone dissenting rubric was right and the two-rubric majority was wrong: npm run validate refused exactly the five records consistency had rejected, no more and no fewer. A 2-of-3 threshold published five claims that a deterministic gate then had to catch. That is the ordering working as designed, and it is also a warning about what a majority is worth here.
The panel is three instances of one model family. Their independence is of reasoning, not of training. A source that is itself wrong can carry all three, and no gate in this pipeline catches a well-formed claim that is simply untrue.
gate-evidence does not verify quote fidelity and says so in its own header comment; a deliberately corrupted quote passes it at exit 0, which was measured rather than assumed. Every verbatim guarantee in this run rests on the run's own byte-level check against the fetched pages, not on any gate.
This run's own scout minted 6 colliding claim ids because its id template omitted the collection, and a family and a release may share an id. gate-evidence caught it. The run's own quote checker had passed all 528 claims without noticing, because it does not test id uniqueness - a gap in this run's verification, reported rather than quietly patched.
9 fetch(es) returned HTTP 403 and 3 returned a body too thin to read. Those creators' channels went dark for this run; the 403s were reported rather than scraped around, and no user agent was spoofed.
A second consecutive tranche is now blocked by the home route critical-path ceiling. One withheld tranche is a budget; two in a row is a structural signal that the dataset cannot keep growing against a fixed ceiling. Raising a ceiling is out of the auto-merging class and is a human decision, so the run withholds and reports instead.

Follow-ups — proposed, not fixed

The claim-id minter in this run's scout should be collection-qualified. It builds `<targetId>-reverify-<date>`, which collides whenever a family and a release share an id.
The run's quote checker should assert claim-id uniqueness. It verified every quote against fetched bytes and still passed a bundle carrying duplicate ids.
Two consecutive refresh runs have now withheld a new-release tranche on the home route critical-path ceiling. Someone should decide whether the ceiling, the home route payload, or the tranche size is the thing to change.
A panel whose rubrics return identical vectors is worth investigating on its own terms: either the rubrics overlap more than intended, or the prompts do.

2026-09-18-1d4a06

Full sweep of all 44 creators, 406 edits published, the Gemini 3.8 Flash tranche withheld for a fifth run

Scope requested: All 44 creators, every catalogued discovery channel, with no creator narrowed out. The dataset recorded a uniform lastScouted of 2026-09-14 across all 44, so there was no partial sweep to resume and found.unswept is honestly empty rather than empty for want of looking. 296 page fetches were attempted and 288 returned content; the eight that did not are recorded as degraded channels, because those creators were looked at and the channel failed, which is a different state from a creator nobody reached.PublishedSep 18, 2026406 edits posted · 76 items withheld

Every catalogued creator was scouted from primary sources in one pass. 482 claims were proposed from 288 pages fetched and hashed in this run; a three-rubric panel cast 1,446 verdicts and accepted 416 at the per-profile threshold. 406 reached the dataset. Eight accepted claims were held before apply and are published as withheld findings rather than dropped: five verifiedAt advances that would have left a model-fit statement older than the fact it rests on, the two-record Gemini 3.8 Flash tranche, which the home route critical- path ceiling refuses for the fifth consecutive run, and one verifiedAt advance held because a gate fixture addresses releases[0] positionally and is falsified by re-dating whichever record sits there. The one substantive addition is OpenAI GPT Live 1, published as a new family, a new release and its documentation source.

PreflightRanClean tree at eb7a4e961c, no refresh pull request left open from a previous run. Both invocation forms were probed for every tool rather than either being assumed: bare drydock, npm and npx are all refused by the PowerShell execution policy on this host, while drydock.cmd reports 0.1.0 and npm.cmd reports 11.9.0. That is installed and blocked, not absent, and the .cmd form was used throughout. The execution policy was not changed. npm ci was used for install; the lockfile was verified unmodified afterwards.
ScoutRan296 pages attempted across 44 creators, 288 returned content. Every quote was sliced verbatim out of bytes fetched and hashed in this run, each carrying a sha256 and a fetch date; no quote was recalled and no search snippet was treated as evidence. A verbatim check re-read all 488 evidence quotes against the saved page bytes and matched 488 of 488. 482 claims proposed: 475 change, 5 add, 2 conflict.
ReviewRanThree independent rubrics per claim, each reviewer blind to the other two and to the scout. 1,446 verdicts over 482 claims, 2-of-3 for the 7 reviewed-profile creators and unanimous 3-of-3 for the 37 long-tail creators; the threshold was read from each bundle's policy field and never inferred. By rubric: provenance 444 accept / 38 reject, consistency 482 / 0, editorial 409 / 73. 416 claims accepted, 66 rejected. The chair cast no vote and reversed none. Mid-stage the provenance panel was found to be running a broken evidence lookup; it was repaired and those four reviewers re-ran on a outcome-neutral instruction, which is recorded in full under caveats.
GatesRanFive deterministic gates, all required, all read by exit code from an unpiped invocation. gate-source-approval and gate-evidence passed on 44 of 44 bundles; gate-scope, gate-dataset and gate-reversals passed on the applied tree. gate-evidence initially refused 21 bundles, correctly: a rejected claim was still carrying a data-changing kind into the gate. The fix was to withdraw rejected claims from the gate bundle and publish them as withheld, not to relax the gate. No gate was skipped, no threshold lowered and no override used.
PublishRanDataset committed before this entry, so the entry could name its own pull request. npm run validate passes on the committed tree: 139 test files, 3,189 tests, exit 0. The pull request auto-merges on a green web-ci; it was not merged by hand, no branch protection was bypassed and nothing was pushed to main directly.

What was found

Scouts: 44
Pages fetched and hashed: 296
Claims proposed: 482

Claims per creator bundle, with the review threshold its profile set
Creator	Policy	Threshold	Claims
01-ai	long-tail	3-of-3	7
ai-singapore	long-tail	3-of-3	4
ai2	long-tail	3-of-3	8
ai21-labs	long-tail	3-of-3	8
aleph-alpha	long-tail	3-of-3	4
alibaba-cloud	pilot	2-of-3	17
amazon	pilot	2-of-3	13
anthropic	pilot	2-of-3	35
apple	long-tail	3-of-3	10
baidu	long-tail	3-of-3	4
bytedance-seed	long-tail	3-of-3	5
cohere	long-tail	3-of-3	10
databricks	long-tail	3-of-3	4
deepseek	long-tail	3-of-3	9
eleutherai	long-tail	3-of-3	9
google-deepmind	pilot	2-of-3	14
hugging-face	long-tail	3-of-3	31
ibm	long-tail	3-of-3	18
kyutai	long-tail	3-of-3	4
lelapa-ai	long-tail	3-of-3	5
lg-ai-research	long-tail	3-of-3	8
liquid-ai	long-tail	3-of-3	4
maritaca-ai	long-tail	3-of-3	5
meta	pilot	2-of-3	39
microsoft	pilot	2-of-3	23
minimax	long-tail	3-of-3	11
mistral-ai	long-tail	3-of-3	18
moonshot-ai	long-tail	3-of-3	7
naver	long-tail	3-of-3	5
nous-research	long-tail	3-of-3	4
nvidia	long-tail	3-of-3	12
openai	pilot	2-of-3	46
openbmb	long-tail	3-of-3	12
reka-ai	long-tail	3-of-3	4
sakana-ai	long-tail	3-of-3	5
sarvam-ai	long-tail	3-of-3	4
snowflake	long-tail	3-of-3	5
stability-ai	long-tail	3-of-3	10
tencent	long-tail	3-of-3	8
tii	long-tail	3-of-3	8
upstage	long-tail	3-of-3	3
xai	long-tail	3-of-3	10
xiaomi	long-tail	3-of-3	4
zhipu-ai	long-tail	3-of-3	8

What those claims proposed to do
Kind	Count	Effect
Change	475	Field values advanced in place on records already in the dataset.
Add	5	New records written into the dataset.
Conflict	2	Recorded that primary sources disagree. Makes no dataset edit by design.

Not covered

Benchmark results and usage observations were not scouted this run. The sweep covered creators, families, releases and their sources only.
Four sources returned HTTP 200 with too little text to quote from, so they yielded no claims even though the fetch itself succeeded. They are listed as degraded channels, because a page that answers with nothing readable is not a page that was read.

Degraded discovery channels

openai — openai-com-news: Seven fetches to openai.com returned HTTP 403. Mitigated through the openai-news-rss and developers.openai.com channels, which did return content, so the creator was still covered; the page budget went to 44 against a nominal 40 doing it.
xai — xai-news: One fetch to https://x.ai/news returned HTTP 403. No alternative channel returned the same material.
meta — meta-ai-site: HTTP 200 with 56 characters of extractable text. Treated as incomplete, not as a page that said nothing.
cohere — cohere-docs: HTTP 200 with 388 characters of extractable text, below what any claim here could be quoted from.
zhipu-ai — zhipu-ai-glm-4-5-announcement: HTTP 200 with 3 characters of extractable text.
bytedance — bytedance-seed-site: HTTP 200 with 19 characters of extractable text.

What was evaluated

Reviewers: 3
Verdicts cast: 1,446
Accepted by panel: 416
Rejected by panel: 66

Deterministic gates and required checks — 3 of 3 checks passed. Exit 0 is a pass; exit 2 means the gate could not run and is never treated as one.
Check	Scope	Result
`gate-scope`	working tree against the merge-base	Passchanged 5, empty false, every path a dataset document under web/src/data. An exit 0 on an empty result would mean nothing was measured and is never a licence to publish.
`gate-dataset`	the applied change	Pass{ "dataDir": "C:\\workspace\\copilot-worktrees\\ModelTree\\abdeslam-menacere-refactored-pancake\\web\\src\\data", "today": "2026-09-18", "counts": {
`gate-reversals`	web/src/data	Pass{ "dataDir": "C:\\workspace\\copilot-worktrees\\ModelTree\\abdeslam-menacere-refactored-pancake\\web\\src\\data", "rejectionsRead": 282, "rejectionsNamingARecord": 18,

Applied over a recorded dissent

These met their threshold and were applied. The objection stands on the record and was not overruled.

amazon-nova-models-page-last-checked-2026-09-18Editorial and entity boundariesThe source-date claim is backed by 'industry-leading conversational quality and pricing', a rank-style marketing quote.
amazon-nova-2-bedrock-announcement-last-checked-2026-09-18Editorial and entity boundariesThe recheck quote uses 'industry-leading price performance', so it substantiates the source through leaderboard-style marketing.
amazon-bedrock-product-page-last-checked-2026-09-18Editorial and entity boundariesThe Bedrock product quote tells users to pick the 'best model', which is serving-platform marketing and a leaderboard framing.
amazon-sagemaker-jumpstart-page-last-checked-2026-09-18ProvenanceReject: evidence lookup verifies presence, but the quote is AWS page title/navigation chrome rather than substantive SageMaker JumpStart content.
amazon-titan-embeddings-ga-announcement-last-checked-2026-09-18ProvenanceReject: evidence lookup verifies presence, but the quote is title/navigation chrome, not substantive body text for the Titan Embeddings announcement.
amazon-nova-2-verified-2026-09-18Editorial and entity boundariesThe family verifiedAt update rests on an 'industry-leading price performance' quote, not a neutral family-boundary statement.
amazon-titan-verified-2026-09-18ProvenanceReject: evidence lookup verifies presence, but the same title/navigation quote does not substantively re-verify the Amazon Titan family record.
anthropic-source-models-overview-last-checkedEditorial and entity boundariesThe selected quote uses ?state-of-the-art? marketing language, so the date advance is not supported by editorially clean source-record evidence.
anthropic-source-sonnet-4-5-announcement-last-checkedEditorial and entity boundariesThe quote relies on ?best coding model,? ?strongest model,? and ?best model? superlatives, which the editorial rubric excludes from source-record support.
anthropic-source-sonnet-5-docs-last-checkedEditorial and entity boundariesThe quote describes Sonnet 5 as ?the best combination of speed and intelligence,? a ranking-style claim that is not editorially clean evidence.
anthropic-source-sonnet-5-announcement-last-checkedEditorial and entity boundariesThe supporting page/quote uses ?most agentic? and benchmark comparison framing, so a routine lastCheckedDate claim is being substantiated through rankings.
anthropic-source-sonnet-4-6-announcement-last-checkedEditorial and entity boundariesThe quote?s ?most capable Sonnet model yet? framing is a comparative ranking boast, not clean evidence for a source date advance.
anthropic-source-opus-4-6-announcement-last-checkedEditorial and entity boundariesThe quote calls Opus 4.6 the creator?s ?smartest model? and continues with capability improvements, which is marketing-rank framing under this rubric.
anthropic-claude-5-verified-at-2026-09-18ProvenanceRejected: the verified quote is a page title plus cookie boilerplate and does not substantiate the Claude 5 family record.
anthropic-claude-4-5-verified-at-2026-09-18Editorial and entity boundariesThe verification quote is built around ?best? and ?strongest? model claims, so the family date advance imports leaderboard-style framing.
anthropic-claude-4-6-verified-at-2026-09-18Editorial and entity boundariesThe family verification relies on ?smartest model? superlative language, which is not an editorially acceptable way to substantiate the record.
anthropic-claude-4-8-verified-at-2026-09-18Editorial and entity boundariesThe quote cites improvements across benchmarks, so the verification is framed through benchmark-boast evidence the rubric excludes.
anthropic-claude-mythos-5-verified-at-2026-09-18ProvenanceRejected: the verified quote is a page title plus cookie boilerplate and does not substantiate the Claude Mythos 5 release.
anthropic-claude-sonnet-5-verified-at-2026-09-18Editorial and entity boundariesThe release verification rests on ?best combination of speed and intelligence,? which is ranking prose and not an acceptable editorial framing.
anthropic-claude-opus-4-6-verified-at-2026-09-18Editorial and entity boundariesThe release verification quote uses ?smartest model? superlative language, so it imports a ranking claim into the evidence framing.
google-source-deepmind-about-last-checkedProvenanceRejected: the verified quote is title/navigation text and does not substantively re-verify the About Google DeepMind source.
google-source-gemini-pro-page-last-checkedProvenanceRejected: the verified quote is title/navigation text and does not substantively re-verify the Gemini Pro source.
google-gemini-3-8-flash-docs-source-addProvenanceRejected: the verified quote is only a page title and language/navigation chrome, not substantive evidence for adding the source.
google-gemini-3-8-flash-release-addEditorial and entity boundariesGoogle DeepMind?s profile says canonical release names should be API model ids, but this proposes the marketing name as canonicalName; the evidence also imports ?most intelligent? ranking prose.
meta-llama-4-announcement-lastcheckeddate-verified-2026-09-18ProvenanceThe corrected lookup verifies the quote, but it is only the article/page title repeated next to site navigation, not substantive source content.
meta-muse-spark-announcement-lastcheckeddate-verified-2026-09-18ProvenanceThe corrected lookup verifies the quote, but it is only the article/page title next to site navigation, not substantive source content.
meta-muse-image-video-announcement-lastcheckeddate-verified-2026-09-18ProvenanceThe corrected lookup verifies the quote, but it is only the article/page title next to site navigation, not substantive source content.
meta-llama-4-scout-verifiedat-verified-2026-09-18Editorial and entity boundariesThe supporting quote uses 'industry leading' ranking language for the release recheck; editorial review rejects rank/superlative framing even on date advances.
microsoft-mai-transcribe-2-announcement-lastcheckeddate-verified-2026-09-18Editorial and entity boundariesThe recheck is framed through a fastest/most accurate/cheapest-in-the-world ranking, which is not editorially clean source metadata.
microsoft-mai-image-2-6-announcement-lastcheckeddate-verified-2026-09-18Editorial and entity boundariesThe chosen quote says 'strongest image model yet', a rank-style marketing claim rather than clean source-date evidence.
microsoft-mai-code-1-1-flash-announcement-lastcheckeddate-verified-2026-09-18Editorial and entity boundariesThe title and quote use better/faster/higher-quality comparative marketing to support a source-date update.
microsoft-mai-transcribe-2-verifiedat-verified-2026-09-18Editorial and entity boundariesThe verifiedAt update is substantiated by a fastest/most accurate/cheapest-in-the-world ranking rather than a clean release-boundary quote.
microsoft-mai-image-2-6-verifiedat-verified-2026-09-18Editorial and entity boundariesThe verifiedAt evidence is 'strongest image model yet', which imports a ranking into the review rather than a neutral release fact.
microsoft-mai-code-1-1-flash-verifiedat-verified-2026-09-18Editorial and entity boundariesThe verifiedAt evidence relies on a higher-quality-code comparative boast, not an editorially clean model-release statement.
source-openai-gpt-5-6-cyber-docs-last-checked-2026-09-18Editorial and entity boundariesThe quote frames GPT-5.6 Cyber as an alias for ?our most advanced? cybersecurity models, mixing alias/release framing with a superlative ranking.
source-openai-gpt-image-2-docs-last-checked-2026-09-18Editorial and entity boundariesThe supporting quote calls GPT Image 2 ?state-of-the-art,? a leaderboard-style boast barred even for routine date advances.
source-openai-gpt-6-astra-docs-last-checked-2026-09-18Editorial and entity boundariesThe supporting quote calls GPT-6 Astra ?our most capable model,? presenting a winner rather than a neutral source re-read.
family-openai-gpt-image-verified-2026-09-18Editorial and entity boundariesThe evidence for the family verification relies on a ?state-of-the-art? claim, which is ranking language rather than clean editorial support.
family-openai-gpt-6-verified-2026-09-18Editorial and entity boundariesThe evidence frames GPT-6 Astra as ?our most capable model,? so the family verification imports a superlative ranking.
release-openai-gpt-5-6-cyber-verified-2026-09-18Editorial and entity boundariesThe quote describes an alias for ?our most advanced? cybersecurity models, so the release verification blurs alias versus release and uses ranking language.
release-openai-gpt-image-2-verified-2026-09-18Editorial and entity boundariesThe release verification is supported by ?state-of-the-art? image-generation phrasing, which the rubric excludes from prose support.
release-openai-gpt-6-astra-verified-2026-09-18Editorial and entity boundariesThe release verification rests on ?our most capable model,? a superlative winner claim rather than neutral model evidence.

Posted 406 edits

406 edits across 3 documents, a net change of 3 records.

Dataset documents this run changed
Document	Before	After	What changed
`web/src/data/sources.json`	295	296	217 edits, almost all lastCheckedDate advances on sources re-read and re-hashed this run. The one new record is openai-gpt-live-1-docs, the primary documentation page the GPT Live 1 release record is built on, fetched and hashed this run. It is recorded here rather than in the posted-records list because the /refresh page links releases, families and organizations only, so a source record there would be a link to nowhere.
`web/src/data/families.json`	86	87	76 edits. The new record is the OpenAI GPT Live family, kept distinct from the GPT-5 line rather than folded into it.
`web/src/data/releases.json`	128	129	113 edits. The new record is OpenAI GPT Live 1. Gemini 3.8 Flash is not here: it was accepted by the panel and then withheld by the asset budget, and is listed under withheld rather than silently absent. The GPT-4.1 verifiedAt advance is also absent and listed under withheld, held by a positional gate fixture.

Each document links to the file as this run left it, not as it stands today.

Records added

GPT-Live 1passportreleasesopenai-gpt-live-1New release, the only substantive addition this run. Modalities, context window and lifecycle status each transcribed from a quoted line of the model reference rather than taken from a sibling model.
GPT-Livein the treefamiliesopenai-gpt-liveNew family for that release. Recorded as its own family rather than as a GPT-5 variant, because the creator documents it as a separate line.

Not posted 76 items

Rejected by the review panel

01-ai-yi-repository-last-checked-2026-09-18Proposed for 01-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Reject: evidence lookup verifies presence, but the quote is GitHub title/navigation chrome rather than substantive Yi README content.Blocked by review rubric: provenance
01-ai-yi-1-5-34b-chat-model-card-last-checked-2026-09-18Proposed for 01-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Reject: evidence lookup verifies presence, but the quote is only the Hugging Face title/navigation area, not substantive Yi-1.5-34B-Chat card text.Blocked by review rubric: provenance
01-ai-yi-34b-chat-model-card-last-checked-2026-09-18Proposed for 01-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Reject: evidence lookup verifies presence, but the quote is only the Hugging Face title/navigation area, not substantive Yi-34B-Chat card text.Blocked by review rubric: provenance
01-ai-yi-1-5-34b-chat-verified-2026-09-18Proposed for 01-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Reject: the quote states only that Yi-1.5 series models are open-sourced; it does not name or substantiate the 34B Chat release itself.Blocked by review rubric: provenance
01-ai-yi-34b-chat-verified-2026-09-18Proposed for 01-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Reject: the quote describes chat models collectively but does not name or substantiate Yi-34B-Chat specifically.Blocked by review rubric: provenance
source-ai2-olmo-2-announcement-last-checked-2026-09-18Proposed for ai2 and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The statement repeats ?the best fully open language model to date,? a direct best-in-class ranking in the claim text.Blocked by review rubric: editorial
source-ai2-olmo-2-model-card-last-checked-2026-09-18Proposed for ai2 and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The evidence cites a ?9-point increase in MMLU? and evaluation improvements, so the date advance is substantiated with benchmark framing.Blocked by review rubric: editorial
source-ai2-molmo-announcement-last-checked-2026-09-18Proposed for ai2 and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The quote calls Molmo ?open state-of-the-art,? which is expressly barred ranking language.Blocked by review rubric: editorial
family-ai2-molmo-verified-at-2026-09-18Proposed for ai2 and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The family verification relies on ?state-of-the-art? wording, importing a leaderboard-style claim into the dataset support.Blocked by review rubric: editorial
ai21-labs-jurassic-1-announcement-last-checked-2026-09-18Proposed for ai21-labs and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The source-date claim uses 'largest and most sophisticated ... ever', an explicit superlative/ranking frame.Blocked by review rubric: editorial
ai21-labs-jurassic-1-jumbo-verified-at-2026-09-18Proposed for ai21-labs and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The release verifiedAt evidence repeats 'largest and most sophisticated ... ever', which is a superlative ranking rather than clean release framing.Blocked by review rubric: editorial
anthropic-source-fable-5-1-announcement-last-checkedProposed for anthropic and rejected by the panel: 1 accept(s) against a 2-of-3 pilot threshold. provenance: Rejected: the verified quote is only the article title, navigation labels, and section headings, not substantive source-record evidence. editorial: The source quote frames the page as ?a new performance frontier,? which is performance-boast framing rather than clean evidence for a routine source re-read.Blocked by review rubric: provenancereview rubric: editorial
anthropic-claude-5-1-verified-at-2026-09-18Proposed for anthropic and rejected by the panel: 1 accept(s) against a 2-of-3 pilot threshold. provenance: Rejected: the verified quote is only title/navigation/section-heading text and does not substantiate the Claude 5.1 family. editorial: The family verification quote uses ?new performance frontier? marketing framing rather than a clean family/release boundary statement.Blocked by review rubric: provenancereview rubric: editorial
apple-openelm-3b-instruct-model-card-last-checked-2026-09-18Proposed for apple and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Rejected: the verified quote is only Hugging Face page title/navigation chrome and does not substantively re-verify the model card.Blocked by review rubric: provenance
apple-fastvlm-model-card-last-checked-2026-09-18Proposed for apple and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Rejected: the verified quote is only Hugging Face page title/navigation chrome and does not substantively re-verify the model card.Blocked by review rubric: provenance
apple-openelm-verified-2026-09-18Proposed for apple and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The family verification quote calls OpenELM ?state-of-the-art,? which is leaderboard-style marketing language barred by the editorial rubric.Blocked by review rubric: editorial
source-cohere-models-overview-last-checked-2026-09-18Proposed for cohere and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: The corrected lookup verifies the quote, but the quote is only the page heading and does not substantiate any model-reference content.Blocked by review rubric: provenance
source-cohere-rerank-v3-5-changelog-last-checked-2026-09-18Proposed for cohere and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The quote calls Rerank 3.5 the 'newest and most performant' model, which is winner/ranking language unsuitable as editorial support.Blocked by review rubric: editorial
release-cohere-rerank-v3-5-verified-at-2026-09-18Proposed for cohere and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: The corrected lookup verifies the quote, but it describes Rerank models generically and does not identify or substantiate the rerank-v3.5 release.Blocked by review rubric: provenance
databricks-dbrx-announcement-last-checked-2026-09-18Proposed for databricks and rejected by the panel: 1 accept(s) against a 3-of-3 long-tail threshold. provenance: The corrected lookup verifies the quote, but it is a page title followed by navigation text, not substantive DBRX source content. editorial: The evidence title frames DBRX as a 'State-of-the-Art Open LLM', a superlative/ranking claim unsuitable for an editorially clean source recheck.Blocked by review rubric: provenancereview rubric: editorial
hugging-face-download-stats-docs-last-checked-2026-09-18Proposed for hugging-face and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Reject: evidence lookup verifies presence, but the quote is page title/navigation chrome and does not substantiate the download-stat policy text.Blocked by review rubric: provenance
hugging-face-smollm3-model-card-last-checked-2026-09-18Proposed for hugging-face and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Reject: evidence lookup verifies presence, but the quote is only the Hugging Face model-card title/navigation, not substantive SmolLM3 card content.Blocked by review rubric: provenance
hugging-face-smollm3-blog-last-checked-2026-09-18Proposed for hugging-face and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Reject: evidence lookup verifies presence, but the quote is the blog title/navigation area and not substantive SmolLM3 article text.Blocked by review rubric: provenance
hugging-face-gguf-docs-last-checked-2026-09-18Proposed for hugging-face and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Reject: evidence lookup verifies presence, but the quote is docs title/navigation chrome and does not substantiate the GGUF documentation.Blocked by review rubric: provenance
liquid-lfm2-announcement-last-checked-2026-09-18Proposed for liquid-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The source recheck is supported by 'sets a new standard in quality, speed, and memory efficiency', a best-in-class marketing claim.Blocked by review rubric: editorial
liquid-lfm2-1-2b-model-card-last-checked-2026-09-18Proposed for liquid-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The model-card source recheck is supported by 'sets a new standard', which is rank-style marketing rather than neutral source metadata.Blocked by review rubric: editorial
minimax-m1-announcement-last-checked-2026-09-18Proposed for minimax and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Reject: evidence lookup verifies presence, but the quote is page title/navigation chrome, not substantive MiniMax-M1 announcement content.Blocked by review rubric: provenance
minimax-m1-80k-model-card-last-checked-2026-09-18Proposed for minimax and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Reject: evidence lookup verifies presence, but the quote is only the Hugging Face title/navigation area, not substantive MiniMax-M1-80k card text.Blocked by review rubric: provenance
minimax-m1-40k-model-card-last-checked-2026-09-18Proposed for minimax and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Reject: evidence lookup verifies presence, but the quote is only the Hugging Face title/navigation area, not substantive MiniMax-M1-40k card text.Blocked by review rubric: provenance
minimax-m1-repository-last-checked-2026-09-18Proposed for minimax and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Reject: evidence lookup verifies presence, but the quote is GitHub page header/navigation rather than README or licence text substantiating the repository source.Blocked by review rubric: provenance
minimax-text-01-model-card-last-checked-2026-09-18Proposed for minimax and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Reject: evidence lookup verifies presence, but the quote is only the Hugging Face title/navigation area, not substantive MiniMax-Text-01 card text.Blocked by review rubric: provenance
minimax-m1-40k-verified-2026-09-18Proposed for minimax and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The release recheck is substantiated with SWE-bench scores, so it uses benchmark material instead of editorially clean release evidence.Blocked by review rubric: editorial
minimax-m1-80k-verified-2026-09-18Proposed for minimax and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The release recheck is substantiated with SWE-bench scores, so it uses benchmark material instead of editorially clean release evidence.Blocked by review rubric: editorial
source-mistral-3-announcement-last-checked-2026-09-18Proposed for mistral-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: The corrected lookup verifies the quote, but it is only the page title plus Mistral site chrome, not substantive source content.Blocked by review rubric: provenance
source-mistral-large-3-model-card-last-checked-2026-09-18Proposed for mistral-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The quote frames Mistral Large 3 as 'state-of-the-art', which is leaderboard/superlative language outside this dataset's editorial scope.Blocked by review rubric: editorial
source-mistral-devstral-2-announcement-last-checked-2026-09-18Proposed for mistral-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: The corrected lookup verifies the quote, but it is only the page title plus Mistral site chrome, not substantive source content.Blocked by review rubric: provenance
source-mistral-small-4-announcement-last-checked-2026-09-18Proposed for mistral-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: The corrected lookup verifies the quote, but it is only the page title plus Mistral site chrome, not substantive source content.Blocked by review rubric: provenance
family-mistral-large-3-verified-at-2026-09-18Proposed for mistral-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The family recheck is substantiated by a 'state-of-the-art' model boast, so the framing imports ranking language rather than a clean entity verification.Blocked by review rubric: editorial
release-mistral-large-3-675b-instruct-verified-at-2026-09-18Proposed for mistral-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The release recheck relies on 'state-of-the-art' phrasing, a rank/superlative claim the editorial rubric excludes.Blocked by review rubric: editorial
moonshot-ai-kimi-k2-model-card-last-checked-2026-09-18Proposed for moonshot-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The source recheck relies on 'state-of-the-art' wording, a superlative/ranking frame excluded by the editorial rubric.Blocked by review rubric: editorial
moonshot-ai-kimi-k2-verified-at-2026-09-18Proposed for moonshot-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The family recheck is substantiated by 'state-of-the-art' language, so it carries rank/superlative framing instead of a clean boundary check.Blocked by review rubric: editorial
moonshot-ai-kimi-k2-instruct-verified-at-2026-09-18Proposed for moonshot-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The release recheck uses 'state-of-the-art' wording, which this rubric treats as disallowed rank/superlative phrasing.Blocked by review rubric: editorial
hyperclovax-seed-text-instruct-1-5b-release-verified-2026-09-18Proposed for naver and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The release verification quote substantiates the record through ?competitive performance on major benchmarks,? which is benchmark/ranking framing.Blocked by review rubric: editorial
naver-hyperclova-x-seed-1-5b-model-card-last-checked-2026-09-18Proposed for naver and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The model-card lastCheckedDate claim uses a quote about competitive benchmark performance, not editorially clean source-record evidence.Blocked by review rubric: editorial
openbmb-minicpm5-1b-model-card-last-checked-dateProposed for openbmb and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The quote says the model reaches ?1B-class open-source SOTA,? which imports state-of-the-art ranking language into a date-advance claim.Blocked by review rubric: editorial
openbmb-minicpm5-2b-model-card-last-checked-dateProposed for openbmb and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The quote says the model reaches ?2B-class open-source SOTA,? a leaderboard-style boast unsuitable for this dataset framing.Blocked by review rubric: editorial
openbmb-minicpm-v-4-5-model-card-last-checked-dateProposed for openbmb and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The supporting quote calls MiniCPM-V 4.5 the ?latest and most capable? model, presenting a ranked winner rather than clean source evidence.Blocked by review rubric: editorial
openbmb-minicpm-v-4-6-model-card-last-checked-dateProposed for openbmb and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The quote calls MiniCPM-V 4.6 the ?most edge-deployment-friendly? model to date, which is superlative ranking language.Blocked by review rubric: editorial
openbmb-minicpm-v-4-5-verified-atProposed for openbmb and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The verification relies on ?latest and most capable? wording, so it imports a superlative ranking into the claim.Blocked by review rubric: editorial
openbmb-minicpm-v-4-6-verified-atProposed for openbmb and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The verification relies on ?most edge-deployment-friendly? wording, which is a ranked-fit claim rather than conditional guidance.Blocked by review rubric: editorial
sakana-ai-evollm-jp-model-card-last-checked-2026-09-18Proposed for sakana-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Reject: the verified quote is only the Hugging Face title/navigation chrome and does not substantively re-verify the model-card source.Blocked by review rubric: provenance
sarvam-m-family-verified-2026-09-18Proposed for sarvam-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The quote emphasizes ?exceptional improvements over the base model,? a benchmark-boast framing rather than a clean family verification statement.Blocked by review rubric: editorial
sarvam-m-v1-release-verified-2026-09-18Proposed for sarvam-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The release verification quote publishes percentage benchmark improvements, which the editorial rubric rejects as benchmark/ranking prose.Blocked by review rubric: editorial
sarvam-ai-sarvam-m-model-card-last-checked-2026-09-18Proposed for sarvam-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The source date advance is supported by ?exceptional improvements? performance-boast language instead of clean source-record evidence.Blocked by review rubric: editorial
sarvam-ai-sarvam-m-blog-last-checked-2026-09-18Proposed for sarvam-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The source date advance relies on explicit benchmark improvement percentages, which are outside the editorial scope for source support.Blocked by review rubric: editorial
stability-ai-sd-3-5-announcement-lastcheckeddate-2026-09-18Proposed for stability-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The source quote says ?superior quality? and ?most powerful,? making the date advance rely on winner/ranking language.Blocked by review rubric: editorial
stability-ai-stable-diffusion-3-5-verifiedat-2026-09-18Proposed for stability-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The family verification uses the same ?superior quality? and ?most powerful? quote, which is excluded by the editorial rubric.Blocked by review rubric: editorial
tencent-hunyuan-video-model-card-lastcheckeddate-2026-09-18Proposed for tencent and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The quote says HunyuanVideo is comparable or superior to leading closed-source models, a comparative ranking claim this rubric rejects.Blocked by review rubric: editorial
tencent-hunyuan-video-verifiedat-2026-09-18Proposed for tencent and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The family recheck is supported by a superiority-to-leading-models quote, which imports ranking/winner framing into the record.Blocked by review rubric: editorial
source-tii-falcon-180b-announcement-last-checked-2026-09-18Proposed for tii and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The statement repeats ?World?s Most Powerful Open LLM,? a superlative winner claim rather than neutral source framing.Blocked by review rubric: editorial
xai-grok-4-announcement-last-checked-2026-09-18Proposed for xai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The evidence quote says Grok 4 is ?the most intelligent model in the world,? an explicit ranking claim unsuitable for a date advance.Blocked by review rubric: editorial
xai-grok-4-5-announcement-last-checked-2026-09-18Proposed for xai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The quote calls Grok 4.5 SpaceXAI?s ?smartest model,? which is superlative marketing rather than editorially clean source-record evidence.Blocked by review rubric: editorial
xai-release-notes-last-checked-2026-09-18Proposed for xai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Rejected: the verified quote is only release-notes title/navigation/loading text and does not substantiate release-note content.Blocked by review rubric: provenance
xai-grok-4-verified-2026-09-18Proposed for xai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The family verification relies on ?most intelligent model in the world,? so it imports a ranking claim the dataset should not publish.Blocked by review rubric: editorial
xiaomi-mimo-7b-rl-0530-model-card-last-checked-dateProposed for xiaomi and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: The corrected lookup verifies the quote, but it is only the Hugging Face page title/navigation and not substantive model-card content.Blocked by review rubric: provenance
xiaomi-mimo-7b-rl-0530-verified-atProposed for xiaomi and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. editorial: The quote advances the release recheck through benchmark and 'surpass DeepSeek R1' language, which is comparative ranking framing.Blocked by review rubric: editorial

Accepted by the panel, then dropped

anthropic-claude-4-5-verified-at-2026-09-18Accepted for anthropic at the pilot threshold and then held before apply: applying this verifiedAt advance makes a model-fit statement older than the fact it rests on, which validate.ts refuses; the guidance was never before this panel, so it is held rather than silently re-datedBlocked by review rubric: editorial
anthropic-claude-mythos-5-verified-at-2026-09-18Accepted for anthropic at the pilot threshold and then held before apply: applying this verifiedAt advance makes a model-fit statement older than the fact it rests on, which validate.ts refuses; the guidance was never before this panel, so it is held rather than silently re-datedBlocked by review rubric: provenance
anthropic-claude-haiku-4-5-verified-at-2026-09-18Accepted for anthropic at the pilot threshold and then held before apply: applying this verifiedAt advance makes a model-fit statement older than the fact it rests on, which validate.ts refuses; the guidance was never before this panel, so it is held rather than silently re-dated
meta-llama-4-scout-verifiedat-verified-2026-09-18Accepted for meta at the pilot threshold and then held before apply: applying this verifiedAt advance makes a model-fit statement older than the fact it rests on, which validate.ts refuses; the guidance was never before this panel, so it is held rather than silently re-datedBlocked by review rubric: editorial
release-openai-gpt-5-verified-2026-09-18Accepted for openai at the pilot threshold and then held before apply: applying this verifiedAt advance makes a model-fit statement older than the fact it rests on, which validate.ts refuses; the guidance was never before this panel, so it is held rather than silently re-dated

Sources conflict, so no value changed

snowflake-arctic-release-date-conflictAccepted for snowflake but kind "conflict" does not touch the dataset by design. Snowflake sources disagree on Arctic release timing: the model card says April 24, 2024, while the Arctic repository says April 2023.
stability-ai-svd-intended-use-conflictAccepted for stability-ai but kind "conflict" does not touch the dataset by design. Stability AI sources disagree on Stable Video Diffusion intended use: the announcement says it is not for real-world or commercial applications at that stage, while the model card says it is intended for both non-commercial and commercial usage.

Blocked by policy before it could run

google-gemini-3-8-flash-docs-source-addAccepted for google-deepmind at the pilot threshold and then held before apply: home route critical-path budget refusal; withheld as a tranche per ADR 0015Blocked by web/asset-budgets.json fixedRoutes[home].criticalMaxRaw
google-gemini-3-8-flash-release-addAccepted for google-deepmind at the pilot threshold and then held before apply: home route critical-path budget refusal; withheld as a tranche per ADR 0015Blocked by web/asset-budgets.json fixedRoutes[home].criticalMaxRaw
release-openai-gpt-4-1-2025-04-14-verified-2026-09-18Accepted for openai at the pilot threshold and then held before apply: this record is releases[0], and the self-test at gates.test.mjs:1175 sets releases[0].verifiedAt to the real today and asserts exactly one change; dating it today makes that mutation inert and reddens the required skills-ci check. The fixture is under .github/ and so outside the ADR 0003 class a refresh may edit, leaving withholding the claim as the only remedy available to this run.Blocked by .github/skills/modeltree-gates/scripts/gates.test.mjs:1175 (positional fixture addressing releases[0])

What this run does not prove

Forty-three of the 1,446 reviewer rationales carry a literal question mark where a typographic quotation mark was intended, lost in the reviewers own prose before it was recorded. None of the 488 evidence quotes is affected, and that split is the part that matters: the evidence is the text that was fetched, hashed and checked verbatim against the page, and it round-tripped clean at 488 of 488. The rationales are published as recorded rather than repaired, because a rationale is a reviewers words and silently rewriting one would make the published trail differ from the verdict that was actually cast.
A green run proves the dataset is internally consistent and that every published claim was sourced, quoted and voted on. It does not prove the underlying page is correct, and it does not prove a page that went dark says what it said when it was last read.
The provenance panel ran a broken instrument for its first pass and the failure was systemic, not per-claim. Its evidence lookup matched quotes against saved page bytes with a raw substring test, so any quote straddling a line wrap in the saved text read as "not verbatim"; and it guessed page filenames from the creator and source id, so pages not named that way read as "absent". Both defects fail toward rejection, which is the direction that looks like rigour while destroying sound claims. The instrument was replaced with one that normalises whitespace on both sides and keys pages by the meta.sourceId recorded inside each page's metadata, over an evidence base of 339 distinct source ids and 339 page files, and the four provenance reviewers were re-run on an instruction that named no desired outcome. Semantic rejections survived the repair; the ones resting on the false premise dissolved. Replacing a reviewer's broken instrument is not vote-rigging; re-running a reviewer to change an outcome is, and that was not done.
The editorial rejections carry the chair's fingerprints and should be read as such. The editorial brief this chair wrote extended the standing "no composite score, rank or winner" rule from the recorded value to the selection of the evidence quote, so a claim whose quote happened to carry a benchmark or marketing superlative was rejected even where the recorded value was unobjectionable. That accounts for the bulk of the 73 editorial rejections. The reviewers applied the rubric as written, so the rejections stand and were not re-run; but the rule is the chair's drafting choice and not a settled policy of this repository, and it is named here so a later reader can disagree with it.
The consistency rubric accepted 482 of 482 and rejected nothing. A rubric that returns one answer to everything has not been shown to discriminate, and this one has not been. Its accepts should carry less weight this run than the other two rubrics', and a later run should give it a claim it ought to refuse before trusting the column.
Three instances of the same model family reading the same page share a failure mode. The 2-of-3 and 3-of-3 thresholds buy independence of reasoning, not independence of training, and a source that is itself wrong can carry all three reviewers.
web/asset-budgets.json is in this run's diff. ADR 0015 amends ADR 0003 to admit that file into the auto-merging class for its regenerable measurement fields only, and gate-scope enforces that with a content-aware check rather than by listing the path. Only measurement figures were re-recorded here; no ceiling and no drift tolerance was touched. That was controlled rather than asserted: with the re-record alone gate-scope exits 0 and reports that only regenerable measurement and prose fields changed, and with home.criticalMaxRaw nudged in the same file it exits 1 and names fixedRoutes[0].criticalMaxRaw. The pass is a reading, not an instrument that says yes to everything.
The home route measurement had been left stale by earlier runs: run 2026-09-13-bfee2c recorded that measuredRaw was already 19,971 bytes out of date and published without re-recording it. That accumulated staleness, not this run's 2,249 bytes of content, is what made the drift check bind tighter than the ceiling here. It is re-recorded now.
One accepted claim was held because a gate fixture, not the dataset, refuses it. gates.test.mjs:1175 copies the live web/src/data, sets releases[0].verifiedAt to today and asserts the mutation changed exactly one field; so re-dating whichever record happens to sit at index 0 makes the fixture's own mutation inert and the assertion fails. That record is currently openai-gpt-4-1-2025-04-14. The diagnosis is a two-arm measurement rather than an inference: with that verifiedAt at today the suite is 440 of 441 with that one test failing, and with it left at 2026-09-01 the suite is 441 of 441, same command and same test filter, one field differing. skills-ci is deliberately unfiltered and required, so a red suite would have blocked auto-merge and reddened main. The fixture is under .github and outside the ADR 0003 class, so it was not edited; the claim was withheld instead, which is the same remedy run 2026-09-14-708ff4 used.
This run's own scratch controls carried two stale fixtures of exactly the kind they exist to catch, and both were repaired by deriving the fixture from the tree rather than pinning it. One named a release id that this run's claims then gave the very field the control needed it to lack, and its vacuity guard fired correctly rather than passing silently. The other assumed the branch carried no commit yet. Repairing them found a real defect described in the follow-ups; a control left broken would have found nothing. A third instrument defect was corrected in the same pass: the draft assembler filed the two recorded conflicts as blocked-by-policy, which is the category for an edit a rule refused rather than for a claim that makes no edit by design. They are published as conflict-recorded, so the three genuinely policy-blocked holds are countable as three.

Follow-ups — proposed, not fixed

gates.test.mjs:1175 addresses releases[0] positionally, so a refresh that re-dates whichever record sits at index 0 falsifies the fixture and reddens a required, deliberately unfiltered workflow. Run 2026-09-14-708ff4 met this, worked around it the same way, and asked that the fixture pin a record the dataset is not free to re-date. This is its second consecutive recurrence, and the workaround costs a sound claim every time. The fixture should address a synthetic record it owns, or assert on a field the refresh class cannot touch, rather than on whatever the dataset happens to sort first.
This run's ledger writer kept its own path allow-list and refused web/asset-budgets.json outright, while the repository's gate-scope admits that file under ADR 0015 by a content-aware check. A second implementation of a rule cannot express a content-aware admission and drifts from the authority silently; this one was caught only because its two-arm control was re-run. It now delegates the question to gate-scope instead of re-deciding it. Any tooling that mirrors a gate's scope rule should call the gate rather than restate it.
The same defect exists in the repository, one layer up, and it disarms the guard that exists to prevent #419. gate-ledger.mjs computes the ADR 0003 qualifying class as a literal path list (DATASET_PATHS plus the ledger, line 397), whose comment asserts it is "the qualifying class exactly as gate-scope.mjs computes it". ADR 0015 made gate-scope content-aware, admitting web/asset-budgets.json when only measurement fields moved; gate-ledger was not widened, so it counts that file out of class. Rule 5 - the rule that requires a change able to reach main unattended to record itself on /refresh - fires only when outOfClass is empty, so it does not fire for any run that re-records budgets. That is exactly the run type that must record itself, and gate-ledger will not require it to. Measured on this branch, two arms: with the ledger entry deleted the gate returned exit 0 and failures [] while reporting unattended: false and outOfClass [web/asset-budgets.json]; the repository's own fixture test "Rule 5 - a dataset change confined to the qualifying class must record itself" passes in the same tree, so rule 5 is live and the arms differ. The self-test at gates.test.mjs:8502 cannot catch this: it compares the two literal path sets, and ADR 0015's admission is a conditional branch outside ALLOWED_PATHS, so the sets still match while the computed classes diverge. This run recorded itself regardless, so nothing is missing from the ledger - the guard simply did not ask. gate-ledger should call gate-scope for the in-class decision, on the same rule the entry above states.
postedRecordSchema.collection is typed as nonEmpty, so the ledger schema accepts any collection name, while the /refresh page resolves exactly three - postedRecordLink in web/src/lib/refresh-log-links.ts handles releases, families and organizations and returns resolved: false for everything else, which refresh-log-links.test.ts then fails the whole log over. This run listed its new source record among its posted records, which validated cleanly against the schema and reddened web-ci; the gates do not see it either, because gate-ledger checks the entry's numbers rather than its link targets. The failure therefore surfaces only at the last check before merge, with a message about an unresolved link rather than about an unsupported collection. Narrowing that field to the collections the page can render would move the refusal to the point of authorship and say what is wrong. Recorded rather than fixed here: refresh-log-schema.ts is outside the ADR 0003 qualifying class, so no refresh run may touch it.
The scout proposed featured: true on a new release record. No refresh may do that: featured-policy.test.ts pins FEATURED_ROSTER as an exact set by identity, and that roster lives in a .ts test file which is outside the ADR 0003 qualifying class, so a refresh can add the flag but can never add the roster entry that makes it legal. The scout brief should exclude featured from proposable fields, and the failure should say so rather than reporting a count mismatch.
Five verifiedAt advances were accepted and then held because model-fit-statements.json carries guidance dated earlier than the facts those advances would create, which validate.ts refuses. The guidance was never in front of the panel, so re-dating it here would have published a re-verification nobody performed. The scout should propose a record's verifiedAt advance together with re-verification of any fit statement resting on it, so the pair is reviewed as one thing or neither moves.
The Gemini 3.8 Flash tranche has now been withheld by the same home route ceiling in four consecutive runs before this one and in this one. It measures 9,748 bytes against 1,640 bytes of headroom on a clean tree, so no amount of unrelated trimming elsewhere in the dataset will admit it. This is a standing decision being re-taken by default rather than decided: either the home route should stop rendering every release in its critical path, or the ceiling should be revisited deliberately by a human under ADR 0015, which this pipeline may not do.
gate-reversals passed while reporting that it could not tell, for most of this run's 66 recorded rejections, whether the rejected value is present in the dataset, because their detail text names no record. That is a limit of the gate rather than a finding that nothing was reversed.

2026-09-14-708ff4

Every creator re-verified; 465 dates advanced and nothing else

Scope requested: All 44 creators, every catalogued discovery channel. 336 page fetches attempted, 326 returned content. The three creators whose channels went dark are recorded as degraded channels rather than as unswept creators, because they were looked at.PublishedSep 14, 2026465 edits posted · 22 items withheld

A full sweep of all 44 creators. 487 claims were proposed, 475 met their threshold, and 465 reached the dataset as verification-date advances across sources.json, releases.json and families.json. Every value is a 10-character ISO date replacing a 10-character ISO date, so the change is 465 insertions against 465 deletions and a net byte delta of zero, and it moves no asset budget. No fact changed: the run re-read what the dataset already asserts and recorded the date it did so. The 22 claims that did not reach the dataset are named individually under withheld, each with the rubric, gate or policy that held it back.

PreflightRanClean tree at 32ae0639ab, no refresh pull request left open from a previous run. Both invocation forms probed for every tool rather than either being assumed: bare drydock, npm and npx are all refused by the PowerShell execution policy on this host, while drydock.cmd reports 0.1.0 and npm.cmd reports 11.9.0. That is installed and blocked, not absent, and the .cmd form was used throughout. The execution policy was not changed.
ScoutRan336 pages attempted across 44 creators, 326 returned content after retries. Every quote was sliced verbatim out of bytes fetched and hashed in this run, each carrying a sha256 and a fetch date; no quote was recalled and no search snippet was treated as evidence. 487 claims proposed, 481 change and 6 add.
ReviewRan21 review agents across 7 packets, three independent rubrics per claim (provenance, consistency, editorial), blind to each other and to the scout. 1,461 verdicts cast over 487 claims. 2-of-3 for the 7 reviewed-profile creators and unanimous 3-of-3 for the 37 long-tail creators. No threshold was lowered and no reviewer was re-run to change an outcome.
GatesRangate-source-approval and gate-evidence ran pre-apply on a clean tree, 44 of 44 bundles at exit 0 each. After applying, gate-scope, gate-dataset (642 records), gate-reversals and gate-ledger all exited 0, and npm run validate passed 3,189 tests with 0 astro-check diagnostics. ci-preflight selected web-ci, skills-ci and source-link-health-tests and exited 0. An earlier ci-preflight run exited 1 on skills-ci; the cause was measured rather than assumed and one claim was withheld, not a test edited.
PublishRanTwo commits on one branch and one pull request: the dataset change first, then this ledger entry, which cannot name its own pull request until that pull request exists. The repository squash-merges, so the two arrive on main as one commit and revert as one. GitHub performs the merge once web-ci is green; the agent does not. No --admin, no --force, no skipped gate and no direct push to main.
DeployNot runRecorded before the merge, so the Pages deploy had not run when this entry was written. Confirmed separately against the full 40-character merge commit afterwards, with a two-armed negative control. Not-run is recorded as itself rather than rounded into a pass.

What was found

Scouts: 44
Pages fetched and hashed: 336
Claims proposed: 487

Claims per creator bundle, with the review threshold its profile set
Creator	Policy	Threshold	Claims
01-ai	long-tail	3-of-3	7
ai-singapore	long-tail	3-of-3	4
ai2	long-tail	3-of-3	8
ai21-labs	long-tail	3-of-3	6
aleph-alpha	long-tail	3-of-3	4
alibaba-cloud	pilot	2-of-3	17
amazon	pilot	2-of-3	13
anthropic	pilot	2-of-3	28
apple	long-tail	3-of-3	10
baidu	long-tail	3-of-3	4
bytedance-seed	long-tail	3-of-3	5
cohere	long-tail	3-of-3	10
databricks	long-tail	3-of-3	4
deepseek	long-tail	3-of-3	10
eleutherai	long-tail	3-of-3	9
google-deepmind	pilot	2-of-3	31
hugging-face	long-tail	3-of-3	31
ibm	long-tail	3-of-3	16
kyutai	long-tail	3-of-3	4
lelapa-ai	long-tail	3-of-3	5
lg-ai-research	long-tail	3-of-3	12
liquid-ai	long-tail	3-of-3	4
maritaca-ai	long-tail	3-of-3	5
meta	pilot	2-of-3	39
microsoft	pilot	2-of-3	23
minimax	long-tail	3-of-3	11
mistral-ai	long-tail	3-of-3	19
moonshot-ai	long-tail	3-of-3	7
naver	long-tail	3-of-3	5
nous-research	long-tail	3-of-3	4
nvidia	long-tail	3-of-3	12
openai	pilot	2-of-3	41
openbmb	long-tail	3-of-3	12
reka-ai	long-tail	3-of-3	4
sakana-ai	long-tail	3-of-3	5
sarvam-ai	long-tail	3-of-3	4
snowflake	long-tail	3-of-3	4
stability-ai	long-tail	3-of-3	9
tencent	long-tail	3-of-3	8
tii	long-tail	3-of-3	8
upstage	long-tail	3-of-3	3
xai	long-tail	3-of-3	10
xiaomi	long-tail	3-of-3	4
zhipu-ai	long-tail	3-of-3	8

What those claims proposed to do
Kind	Count	Effect
Change	481	Advances a verifiedAt or lastCheckedDate on a record the run re-read and found unchanged.
Add	6	Proposes a new source or release record. None reached the dataset this run.

Not covered

Creators absent from the catalogue are outside every run by construction; this run adds none.
Nothing was fetched at runtime and no live API was polled; every fact rests on a page fetched and hashed during the run.

Degraded discovery channels

openai — official-announcement: openai.com returned HTTP 403 on 8 URLs, dark for a fifth consecutive run. No user-agent was spoofed and no cache was substituted for the source.
x-ai — official-announcement: x.ai returned HTTP 403.
z-ai — official-announcement: z.ai returned HTTP 404.

What was evaluated

Reviewers: 3
Verdicts cast: 1,461
Accepted by panel: 475
Rejected by panel: 12

Deterministic gates and required checks — 8 of 8 checks passed. Exit 0 is a pass; exit 2 means the gate could not run and is never treated as one.
Check	Scope	Result
`gate-source-approval`	all 44 bundles, pre-apply, clean tree	PassEvery citation rests on an approved origin: 51 origins drawn from 295 dataset source records and 20 updater profile catalogues, anchored at 32ae0639ab. 5 citations were refused and their claims dropped.
`gate-evidence`	all 44 bundles	PassEvery published claim carries three rubric verdicts, one each of provenance, consistency and editorial, with a vote and a non-empty rationale.
`gate-scope`	committed HEAD against the anchor	Passchanged 3, empty false, every path a dataset document under web/src/data. An exit 0 on an empty result would mean nothing was measured and is never a licence to publish.
`gate-dataset`	the applied change, 642 records	PassReferential integrity, date ordering and schema shape across families, releases and sources.
`gate-reversals`	web/src/data	PassNo previously rejected claim was reinstated. The gate reports its own blind spot: 186 of 204 rejections name no record, so it cannot tell whether those were reversed. Not checked is not passed.
`gate-ledger`	this entry against the diff it describes	PassDeclared documents match the changed documents, record counts match at both ends, the run id is new and added once, and no prior entry was touched or reordered. transcription false, because dataset documents did change.
`npm run validate`	web/, published change	Pass3,189 tests across 139 files, 0 astro-check errors, warnings or hints. Read unpiped: an earlier piped read reported exit 1 for this same passing run, which was a pipeline artefact and not a result.
`ci-preflight`	web-ci, skills-ci, source-link-health-tests	PassSelected 3 of 7 check groups from the branch diff. An earlier run exited 1 on skills-ci; the cause was measured to this run rather than assumed, and one claim was withheld to resolve it. Exit 2 would never have been a pass.

Applied over a recorded dissent

These met their threshold and were applied. The objection stands on the record and was not overruled.

add-release-google-gemini-3-8-flashEditorial and entity boundariesObjected to the superlative "most intelligent Flash model". The claim met its 2-of-3 threshold and was accepted with the dissent recorded rather than the wording being edited to satisfy the objection or the reviewer re-run. The summary attributes the superlative to Google rather than asserting it in ModelTree's own voice. The claim is withheld regardless, on the byte ceiling.

Posted 465 edits

465 edits across 3 documents, a net change of 0 records.

Dataset documents this run changed
Document	Before	After	What changed
`web/src/data/sources.json`	295	295	275 lastCheckedDate advances. No record added or removed; only the date each source was last re-read moved.
`web/src/data/releases.json`	128	128	116 verifiedAt advances. A 117th was accepted and withheld, so this document understates when one release was last verified rather than overstating it.
`web/src/data/families.json`	86	86	74 verifiedAt advances. No record added or removed.

Each document links to the file as this run left it, not as it stands today.

Not posted 22 items

Rejected by the review panel

redate-qwen3-8-flash-next-licenseProposed for alibaba-cloud and rejected by the panel: 1 accept(s) against a 2-of-3 pilot threshold. provenance: The record is titled 'Qwen Community License 1.0 (Qwen3.8-Flash-Next)' and is cited as the licence text at huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/main/LICENSE. I verified the actual Qwen Community License text is present on the fetched page (body 00d6143767f11890.body offset 36586, 'Qwen Community License 1.0 Copyright (c) 2026 Qwen Permission is hereby granted...'), so the material is still there — but the attached quote is chat-template Jinja code about reasoning-effort levels ('Supported types are xhigh (default), medium, and low'), which has nothing to do with a licence. Read alone, this quote does not show the page still carries the licence text it is cited for. editorial: Record is a LICENSE document (Qwen Community License 1.0) at /blob/main/LICENSE. Attached quote is Jinja chat-template code about reasoning_effort variables ('xhigh', 'medium', 'low'), which is noBlocked by review rubric: provenancereview rubric: editorial
reverify-release-anthropic-claude-fable-5Proposed for anthropic and rejected by the panel: 1 accept(s) against a 2-of-3 pilot threshold. provenance: Entity mismatch. The record's target is release 'Claude Fable 5' (legacy), but the quote 'Use Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5 at higher effort still fall short. All current' names only Claude Fable 5.1 — a distinct sibling release. I verified the phrase 'Fable 5 ' (with trailing space) is absent from the fetched models overview body d9e8980f3bd02b6d.body; only 'Claude Fable 5.1' appears. Read alone, the quote does not re-verify Claude Fable 5. editorial: Record is the Claude Fable 5 release (a distinct, legacy release from Fable 5.1). Attached quote from the models-overview page mentions only 'Claude Fable 5.1' and 'Claude Opus 5' — sibling releases in different families (anthropic-claude-5-1, anthropic-claude-5) — and does not name Fable 5 at all. A quote about Fable 5.1 does not re-verify the Fable Blocked by review rubric: provenancereview rubric: editorial
redate-osi-license-mitProposed for deepseek and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: The record is 'The MIT License' at opensource.org/license/mit; it is cited as the MIT licence text. I verified the actual MIT licence text IS present on the fetched page (body 04c5c4d77b36d353.body, 'Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files'), but the attached quote 'Donate to the OSI The OSI is the authority that defines Open Source, recognized globally by individuals, companies, and public institutions' is a site-wide OSI donation/mission banner appearing after the licence body. Read alone, this quote shows only that an OSI page rendered; it does not demonstrate the MIT licence material is still present.Blocked by review rubric: provenance
redate-deepseek-v4-flash-model-cardProposed for deepseek and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Entity mismatch. The record's target is source deepseek-v4-flash-model-card (URL huggingface.co/deepseek-ai/DeepSeek-V4-Flash), but the attached quote 'DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2' names DeepSeek-V4-Pro, a distinct sibling release. Read alone, this quote does not demonstrate the V4-Flash page still carries V4-Flash material even though the page URL is correct.Blocked by review rubric: provenance
reverify-family-google-gemini-3Proposed for google-deepmind and rejected by the panel: 1 accept(s) against a 2-of-3 pilot threshold. provenance: Quote is a generic definition of "deprecation" ("A deprecated model is one that we no longer recommend...") — does not mention "Gemini 3" as a family. Grep of the saved deprecations page confirms it does contain a "Gemini 3 models" section, but the ATTACHED quote alone does not carry that material. editorial: Family record is Gemini 3; the quote is only the generic definition of 'deprecation' from the API deprecations page and does not name Gemini 3 anywhere. The quote does not speak to the record's own scope (the Gemini 3 family), so it fails to re-verify that this family is still named on the cited source.Blocked by review rubric: provenancereview rubric: editorial
reverify-family-google-gemini-2-5Proposed for google-deepmind and rejected by the panel: 1 accept(s) against a 2-of-3 pilot threshold. provenance: Quote is the same generic "deprecation" definition, with no mention of "Gemini 2.5". The rubric asks whether the quote read alone shows the page carries material about the family — this quote does not name the family at all. editorial: Family record is Gemini 2.5; the quote is the generic 'deprecation' definition and does not name Gemini 2.5 anywhere. Nothing in the shown quote demonstrates the Gemini 2.5 family is named on the cited page — quote is entity-orthogonal to the record.Blocked by review rubric: provenancereview rubric: editorial
reverify-release-kyutai-moshiko-pytorch-bf16Proposed for kyutai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: Quote 'Bias, Risks, and Limitations The model has been trained with a few safeguards to try to limit potential toxic usages, however our toxicity analysis shows that it behaves in the middle of existing models with respect to textual generation' does not name 'moshiko-pytorch-bf16', 'moshiko', 'Moshi', or 'bf16'. Read alone it identifies no model — it is generic model-card boilerplate that could appear on any model page. The sibling source claim on the same URL quotes 'Moshi' and would have carried this release too; this quote does not.Blocked by review rubric: provenance
add-family-lg-ai-research-exaone-4-5Proposed for lg-ai-research and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: firstReleaseDate='2026-04-09' is not supported by any attached evidence quote. The two attached quotes are 'We release EXAONE 4.5, the first open-weight vision language model with 33B params' and 'Integrating a dedicated visual encoder into the existing EXAONE 4.0 framework, we expand the model's capability toward multimodality.' — neither states the date. The full News-list line on the page is '- 2026.04.09 : 🚩 We release EXAONE 4.5, the first open-weight vision language model with 33B params' (verified via ctx.mjs), but the scout truncated the evidence quote to the tail only, so the specific field firstReleaseDate has no quote that states it read alone. This is exactly the missing-quote-for-a-field failure the rubric's worked reject example (openai-gpt-5-3-codex-release-add: 'The page states the modalities, but the claim attached no quote for them. The remedy is to attach Blocked by review rubric: provenance
add-release-lg-ai-research-exaone-4-5-33bProposed for lg-ai-research and rejected by the panel: 0 accept(s) against a 3-of-3 long-tail threshold. provenance: Two independent provenance failures. (1) releaseDate='2026-04-09' is not stated by any of the three attached evidence quotes ('We release EXAONE 4.5, the first open-weight vision language model with 33B params', 'EXAONE 4.5 features 33 billion parameters in total, including 1.2 billion parameters from the vision encoder.', 'Context Length: 262,144 tokens'). The date is in the News list on the page but the scout truncated the quote — same reject-example shape as above. (2) accessType='open-weights' (plural) is not a member of the schema's accessType vocabulary in web/src/data/schema.ts, which is exactly {proprietary-hosted, open-weight, source-available, both, unknown}. Under the vocabulary section of the rubric a mapping is admitted only when 'one member fits' — no schema member 'open-weights' exists, so no member fits and the recording step fails. The quote 'the first open-wBlocked by review rubric: provenancereview rubric: consistencyreview rubric: editorial
reverify-family-openai-gpt-imageProposed for openai and rejected by the panel: 1 accept(s) against a 2-of-3 pilot threshold. provenance: Entity mismatch. The quote from the OpenAI news RSS feed is a list of URL slugs ("new-chatgpt-images-is-here", "staying-ahead-in-the-age-of-ai", "bny") — the rubric explicitly treats a list of unrelated links as insufficient to re-verify an entity. The only arguably related slug is "chatgpt-images", which is a ChatGPT UI feature and not literally the GPT-Image API family; making the connection requires an inference the quote does not contain. Better evidence (the RSS also carries "introducing-chatgpt-images-2-0" from Apr 21 2026 matching firstReleaseDate) was available but not quoted. editorial: Family record is GPT-Image (the model family). The evidence quote from the OpenAI News RSS feed only shows 'new-chatgpt-images-is-here', which is a ChatGPT product feature announcement. ChatGPT is a product; GPT-Image is a model family — different entities. The quoted content does notBlocked by review rubric: provenancereview rubric: editorial
reverify-release-stability-ai-svd-img2vid-xtProposed for stability-ai and rejected by the panel: 2 accept(s) against a 3-of-3 long-tail threshold. provenance: The provided quote — "For details on the user study, we refer to the research paper Uses Direct Use The model is intended for both non-commercial and commercial usage" — does not name Stable Video Diffusion or img2vid-xt at all; "the model" is a generic pronoun that could be from any model card. Read alone as the rubric requires, the quote does not identify the release entity. Better identifying quotes exist on the saved page (e.g. "Stable Video Diffusion Image-to-Video Model Card" at offset 1601) but were not selected.Blocked by review rubric: provenance
reverify-release-xiaomi-mimo-7b-rl-0530Proposed for xiaomi and rejected by the panel: 1 accept(s) against a 3-of-3 long-tail threshold. provenance: Quote is a generic sentence "achieving uniform and simultaneous improvements ... within a small model is challenging" — makes no mention of "MiMo", "Xiaomi", "7B-RL", or the "0530" checkpoint. Read alone, this quote does not tie the page to this specific release. editorial: Release record canonicalName 'MiMo-7B-RL-0530'. The quote ('Moreover, it was widely considered that achieving uniform and simultaneous improvements in both mathematical and code capabilities within a small model is challenging.') does not name MiMo-7B-RL-0530 anywhere and is a general framing sentence about small models. Nothing in the shown quote demonstrates this specific dated release is still named on the source; entity-scope evidence is absent.Blocked by review rubric: provenancereview rubric: editorial

Accepted by the panel, then dropped

add-source-lgai-exaone-4-5-repositoryAccepted 3/3. Withheld because the records it exists to cite — add-family-lg-ai-research-exaone-4-5 and add-release-lg-ai-research-exaone-4-5-33b — were rejected by the panel. Publishing it alone would leave a source citing an EXAONE 4.5 family and release that are not in the dataset. Schema-legal (validate.ts carries no orphan-source rule) but editorially wrong, and it spends headroom on a route already at 4,476 spare bytes.Blocked by add-family-lg-ai-research-exaone-4-5add-release-lg-ai-research-exaone-4-5-33b
add-source-lgai-exaone-4-5-33b-model-cardAccepted 3/3. Withheld for the same reason as the sibling EXAONE 4.5 source: the family and release it cites were rejected, so it would land as an orphan.Blocked by add-family-lg-ai-research-exaone-4-5add-release-lg-ai-research-exaone-4-5-33b

Source refused by the approval gate

redate-lgai-exaone-4-0-repositoryAccepted by the panel, then refused by gate-source-approval and dropped before anything was applied. was read from https://raw.githubusercontent.com but is filed under source "lgai-exaone-4-0-repository", which is https://github.com; evidence must come from the source it cites The source was not added to sources.json to make the citation resolve; that is the move this gate exists to refuse.Blocked by gate-source-approval
reverify-family-lg-ai-research-exaone-4-0Accepted by the panel, then refused by gate-source-approval and dropped before anything was applied. was read from https://raw.githubusercontent.com but is filed under source "lgai-exaone-4-0-repository", which is https://github.com; evidence must come from the source it cites The source was not added to sources.json to make the citation resolve; that is the move this gate exists to refuse.Blocked by gate-source-approval
redate-tencent-hunyuanimage-3-0-repositoryAccepted by the panel, then refused by gate-source-approval and dropped before anything was applied. was read from https://raw.githubusercontent.com but is filed under source "tencent-hunyuanimage-3-0-repository", which is https://github.com; evidence must come from the source it cites The source was not added to sources.json to make the citation resolve; that is the move this gate exists to refuse.Blocked by gate-source-approval
reverify-release-tencent-hunyuanimage-3-0-standardAccepted by the panel, then refused by gate-source-approval and dropped before anything was applied. was read from https://raw.githubusercontent.com but is filed under source "tencent-hunyuanimage-3-0-repository", which is https://github.com; evidence must come from the source it cites The source was not added to sources.json to make the citation resolve; that is the move this gate exists to refuse.Blocked by gate-source-approval
reverify-family-tencent-hunyuanimage-3-0Accepted by the panel, then refused by gate-source-approval and dropped before anything was applied. was read from https://raw.githubusercontent.com but is filed under source "tencent-hunyuanimage-3-0-repository", which is https://github.com; evidence must come from the source it cites The source was not added to sources.json to make the citation resolve; that is the move this gate exists to refuse.Blocked by gate-source-approval

Blocked by policy before it could run

add-source-google-gemini-3-8-flash-docsAccepted 3/3. Withheld with its release: the home route measures 1,100,524 raw critical bytes against a criticalMaxRaw of 1,105,000, leaving 4,476 spare, and the Gemini 3.8 Flash tranche measures 9,360 bytes (applied build: 1,109,884, over by 4,884). This source exists to cite the release, so it is withheld with it rather than split to fit.Blocked by web/asset-budgets.json home.criticalMaxRaw
add-release-google-gemini-3-8-flashAccepted 2/3 under the pilot bar, with an editorial dissent recorded. Withheld on a measured budget refusal, not on the dissent: applying the tranche builds the home route at 1,109,884 raw critical bytes against a ceiling of 1,105,000 (over by 4,884; 9,360 needed against 4,476 spare). The runway report independently rates route:home as REFUSED at 0 further creators.Blocked by web/asset-budgets.json home.criticalMaxRaw
reverify-release-openai-gpt-4-1-2025-04-14Accepted 2/2 against the openai pilot threshold, and cleared gate-source-approval, gate-evidence, gate-scope and gate-dataset. Withheld because publishing it reddens skills-ci. The fixture at .github/skills/modeltree-gates/scripts/gates.test.mjs:1175 ("a record verified today is accepted rather than read as the future") mutates releases[0] positionally, setting its verifiedAt to the real today and declaring "changed 1 releases.verifiedAt"; its assertTouched guard then fails the run if that mutation changed nothing. releases[0] is this record, so advancing it to 2026-09-14 makes the fixture inert on the day the refresh lands. Measured both ways in one session: with the advance applied, gates.test.mjs exits 1 on that assertion; with the advance held back, it exits 0 and the assertion passes. The repair belongs in the fixture, which is out of class for this run, so the claim is withheld and the fixture fragility is filed as a follow-up. Withholding fails safe: the record understates when it was last verified and never overstates it.Blocked by .github/skills/modeltree-gates/scripts/gates.test.mjs:1175docs/adr/0003-an-agent-gated-data-refresh-may-auto-merge.md

What this run does not prove

A green run proves the dataset is internally consistent and that every published claim was sourced, quoted and voted on. It does not prove the underlying page is correct, and it does not prove a page that went dark says what it said when it was last read.
gate-reversals passed while reporting that 186 of its 204 recorded rejections name no record in their detail, so it could not tell whether those were reversed. That is a limit of the gate, not a finding that they were not.
Three creators had a discovery channel return an error this run: openai.com 403 on 8 URLs, x.ai 403 and z.ai 404. Those creators were scouted, so anything announced only on a dark channel would have been missed without anything here showing it.
One accepted claim was withheld because publishing it reddens skills-ci: the fixture at .github/skills/modeltree-gates/scripts/gates.test.mjs:1175 mutates releases[0] positionally and goes inert when that record is dated today. The dataset now understates when that release was last verified. The repair belongs in the fixture, which is out of class for ADR 0003, so it is filed as a follow-up rather than fixed here.
The Gemini 3.8 Flash release and its source were accepted and withheld for a third consecutive run on the home route byte ceiling: 1,100,524 critical raw bytes against a 1,105,000 ceiling leaves 4,476 spare and the tranche costs 9,360. The dataset is therefore knowingly missing a shipped model, and assets:runway rates route:home REFUSED at 0 more creators.
No human reviewed this change. It reaches main under ADR 0003 on the agent panel and the deterministic gates alone, and GitHub performs the merge once web-ci is green.
This entry was written before the merge, so its deploy stage is not-run rather than a pass. The Pages deploy is confirmed separately against the full merge commit.

Follow-ups — proposed, not fixed

The fixture at .github/skills/modeltree-gates/scripts/gates.test.mjs:1175 addresses releases[0] positionally and sets its verifiedAt to the real today, so any refresh that dates the first release today makes the mutation inert and reddens skills-ci on the day it lands. This run met it and worked around it by withholding a claim; the fixture should pin a record that the dataset is not free to re-date, or assert on a record it inserts itself.
Five claims were refused by gate-source-approval for citing https://raw.githubusercontent.com against a source record filed under https://github.com. Filing GitHub READMEs under https://github.com/<owner>/<repo>/raw/main/README.md keeps the citation on an already-approved origin; the EXAONE 4.5 source claim did exactly that and passed approval.
openai.com has returned 403 to this run for five consecutive runs, so OpenAI release discovery now rests entirely on the developer documentation host. A second approved discovery channel for that creator would remove the single point of failure.

2026-09-13-bfee2c

Gemini 3.8 Flash clears the panel and is refused by the home byte ceiling for the second run running

Scope requested: Every creator in organizations.json plus the long-tail profile: 44 scouted, none unswept. 92 pages attempted across 2 fetch rounds, 88 returning content.PublishedSep 13, 20261 edit posted · 4 items withheld

A full sweep of all 44 creators. Only google-deepmind yielded claims: Gemini 3.8 Flash, GA on 2026-09-02, with its Gemini API model page and its Vertex AI model page as sources, plus a featured-rationale restate on Gemini 3.7 Flash whose current wording calls it the newest Flash tier. The three-rubric panel accepted all five claims and gate-evidence admitted the bundle at exit 0. The deterministic asset budget then refused the tranche: the home route measures 1,100,526 raw critical bytes on clean HEAD against a criticalMaxRaw of 1,105,000, and the three new records take it to 1,110,273 - over by 5,273. A gate cannot be outvoted by a panel majority, so the four records that carry bytes were withheld and only the zero-byte verification advance on the Gemini deprecations page was published. This is the second consecutive run to reach the same wall: 2026-09-11-cd4874 withheld the same records for the same reason.

PreflightRanClean tree, gh authenticated, no refresh pull request from a previous run still open. Drydock probed both forms: the bare name is refused by the execution policy and drydock.cmd reports 0.1.0, so the CLI is installed and blocked rather than absent. npm probed the same way: bare refused, npm.cmd reports 11.9.0.
ScoutRan92 pages attempted, 88 ok, 4 failed. Every quote is byte-verbatim from a fetched page, verified mechanically against the saved body by the bundle builder, which throws rather than emits when a quote is absent. Every page is sha256-hashed on the day. No search snippet was used as evidence. One bundle carried claims; 43 creators were scouted and yielded nothing.
ReviewRanThree sub-agents, one per rubric, on three different models, run in parallel and blind to each other, to the scout reasoning and to the running tally; each saw only the claims, their evidence and the dataset slice they touch. 15 verdicts over 5 claims. The chair neither voted nor broke a tie, and no reviewer was re-run.
GatesRangate-source-approval and gate-evidence ran pre-apply on a clean tree. gate-evidence passed at exit 0 and was controlled in the same session against two mutated bundles - one taken sub-threshold, one with retrieval set to search - which it refused at exit 1 apiece, so the pass is a reading rather than a blind instrument. gate-dataset, gate-reversals, gate-scope and gate-ledger passed on the applied tranche. npm run validate refused the full tranche on the asset budget, which is what withheld it.
PublishRanTwo commits on one branch and one pull request: the dataset change first, then this ledger entry, which cannot be written until the pull request it must name exists. Merged by GitHub under --auto --squash once web-ci was green. No --admin, no --force, no skipped gate and no direct push to main.
DeployNot runRecorded before the merge, so the Pages deploy had not yet run when this entry was written. Confirmed separately after merge; a failed deploy is reverted by pull request rather than left to freeze the published site.

What was found

Scouts: 1
Pages fetched and hashed: 92
Claims proposed: 5

Claims per creator bundle, with the review threshold its profile set
Creator	Policy	Threshold	Claims
openai	pilot	2-of-3	0
anthropic	pilot	2-of-3	0
google-deepmind	pilot	2-of-3	5
meta	pilot	2-of-3	0
xai	long-tail	3-of-3	0
mistral-ai	long-tail	3-of-3	0
deepseek	long-tail	3-of-3	0
alibaba-cloud	pilot	2-of-3	0
microsoft	pilot	2-of-3	0
amazon	pilot	2-of-3	0
cohere	long-tail	3-of-3	0
ai2	long-tail	3-of-3	0
tii	long-tail	3-of-3	0
nvidia	long-tail	3-of-3	0
ai21-labs	long-tail	3-of-3	0
zhipu-ai	long-tail	3-of-3	0
moonshot-ai	long-tail	3-of-3	0
eleutherai	long-tail	3-of-3	0
lg-ai-research	long-tail	3-of-3	0
snowflake	long-tail	3-of-3	0
upstage	long-tail	3-of-3	0
ibm	long-tail	3-of-3	0
baidu	long-tail	3-of-3	0
tencent	long-tail	3-of-3	0
bytedance-seed	long-tail	3-of-3	0
stability-ai	long-tail	3-of-3	0
databricks	long-tail	3-of-3	0
minimax	long-tail	3-of-3	0
apple	long-tail	3-of-3	0
hugging-face	long-tail	3-of-3	0
01-ai	long-tail	3-of-3	0
sakana-ai	long-tail	3-of-3	0
sarvam-ai	long-tail	3-of-3	0
naver	long-tail	3-of-3	0
aleph-alpha	long-tail	3-of-3	0
reka-ai	long-tail	3-of-3	0
nous-research	long-tail	3-of-3	0
liquid-ai	long-tail	3-of-3	0
xiaomi	long-tail	3-of-3	0
ai-singapore	long-tail	3-of-3	0
kyutai	long-tail	3-of-3	0
lelapa-ai	long-tail	3-of-3	0
maritaca-ai	long-tail	3-of-3	0
openbmb	long-tail	3-of-3	0

What those claims proposed to do
Kind	Count	Effect
Add	3	The Gemini 3.8 Flash release and its two documentation sources. All three were accepted by the panel and then withheld by the home critical-path byte ceiling; none reached the dataset.
Change	2	A featured-rationale restate on Gemini 3.7 Flash, withheld with the record it depends on, and a verification-date advance on the Gemini deprecations source, which is the one edit this run published.

Not covered

Pages behind an authentication wall or a client-side rendering shell were not executed; only the served bytes were read.
Benchmark figures and vendor performance claims were not extracted at all this run, by policy: they are contested, versioned and not a fact about a release.
Creators other than google-deepmind were swept for novel tokens against the committed dataset; a change that introduces no new token would not have been seen by that filter.

Degraded discovery channels

openai — official-announcement: openai.com/news/ returned HTTP 403 with a 9,908 byte body carrying no article text. This channel has now been dark across four consecutive runs. Reported rather than worked around: no user-agent was spoofed and no cache was substituted for the source.
xai — official-announcement: x.ai/news returned HTTP 403, so xai has no readable announcement channel this run.
zhipu-ai — official-announcement: z.ai/blog returned HTTP 404. The channel recorded in the catalogue no longer resolves.

What was evaluated

Reviewers: 3
Verdicts cast: 15
Accepted by panel: 5
Rejected by panel: 0

Deterministic gates and required checks — 7 of 8 checks passed. Exit 0 is a pass; exit 2 means the gate could not run and is never treated as one.
Check	Scope	Exit	Result
`gate-source-approval`	google-deepmind bundle, pre-apply on a clean tree	0	PassAnchored on merge-base 04116fc43d20d49dc58acbdeb2f66a53ae351ef1. Both proposed origins are already stood behind by the committed dataset. This run approved no new publisher.
`gate-evidence`	google-deepmind bundle	0	Pass5 claims admissible, 5 applicable, 0 failures, policy derived as pilot from the reviewed-profile set on disk rather than read from the bundle. Controlled in the same session: a sub-threshold mutation and a retrieval=search mutation were each refused at exit 1, so the exit 0 discriminates.
`gate-dataset`	applied change	0	PassZero failures. Referential integrity and entity boundaries hold across the changed document.
`gate-reversals`	applied change	0	PassPasses, and the pass is narrower than it looks: the 2026-09-06 rejection of these same Gemini 3.8 records sits in this gate's acknowledged blind spot, so it was not checked rather than checked and cleared. Disclosed here because not-looked and looked-and-found-nothing are not the same reading.
`npm run validate`not required	web/, full tranche	1	FailRecorded as a gate that did its job. 3186 of 3189 tests passed; the 3 failures were the home critical-path budget at 1,110,273 over a 1,105,000 ceiling, and the measuredRaw drift that follows from it. This is what withheld the tranche. No ceiling was raised and no drift allowance was widened.
`npm run validate`	web/, published change	0	PassThe published edit is a two-character date advance and adds no bytes to any route. Exit code read from an unpiped invocation.
`gate-scope`	committed HEAD and working tree	0	PassOnly web/src/data documents moved. web/asset-budgets.json was deliberately not touched: gate-scope would permit regenerating its measuredRaw field, but this run was instructed to confine itself to the dataset, and regenerating it would not have unblocked the tranche, which is refused by a ceiling rather than by the drift guard.
`gate-ledger`	this entry against the diff it describes	0	Passtranscription false. The entry describes the diff that shipped, including what did not ship and why.

Applied over a recorded dissent

These met their threshold and were applied. The objection stands on the record and was not overruled.

google-gemini-3-8-flash-release-addEditorial and entity boundariesRejected while provenance and consistency accepted, so the claim cleared the 2-of-3 pilot bar at 2. The dissent is that the proposed summary calls the release the most intelligent Flash tier of the Gemini 3 family, which is ranking language in the dataset's own voice and cannot stay true once a later Flash model ships. The chair did not vote, did not break the tie and did not reword the claim after the panel had read it, because publishing amended text would publish something no reviewer reviewed. The record was withheld on bytes before the wording could reach the dataset, and the wording is carried to a follow-up.

Posted 1 edit

1 edit across 1 document, a net change of 0 records.

Dataset documents this run changed
Document	Before	After	What changed
`web/src/data/sources.json`	295	295	One field advanced on google-gemini-deprecations: lastCheckedDate 2026-09-11 to 2026-09-13, attested by a page refetched and hashed on the day, which still carries the gemini-3.8-flash row with its September 2, 2026 release date and no announced shutdown date. Accepted unanimously against a 2-of-3 pilot threshold. No record was added or removed, so the document count is unchanged.

Each document links to the file as this run left it, not as it stands today.

Not posted 4 items

Accepted by the panel, then dropped

google-gemini-3-7-flash-featured-rationale-restateAccepted unanimously, then dropped because its replacement text names Gemini 3.8 Flash as the model 3.7 Flash is the documented efficiency fallback from, and that record was withheld. The panel read this claim as part of a tranche in which 3.8 lands; publishing it alone would publish a configuration nobody reviewed. The wording it replaces - calling 3.7 Flash the newest generally available Flash tier - remains true of the dataset as it now stands, though it is already false of the world.Blocked by google-gemini-3-8-flashweb/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md

Blocked by policy before it could run

google-gemini-3-8-flashAccepted 2-of-3 by the panel and admitted by gate-evidence, then refused by the home route critical-path budget. Measured on this tree: clean HEAD 1,100,526 raw critical bytes, the tranche 1,110,273, ceiling 1,105,000, so it is over by 5,273. Real headroom on clean HEAD is 4,474 bytes, not the 24,445 that asset-budgets.json implies, because its recorded measuredRaw of 1,080,555 is 19,971 bytes stale. No ceiling was raised to admit this record.Blocked by web/asset-budgets.json criticalMaxRawweb/tests/build/asset-budgets.test.tsdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md
google-gemini-3-8-flash-docsWithheld with the release it exists to cite. A source registered for a record that is not in the dataset would be an unreferenced entry, so the pair is kept together rather than split to fit.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md
google-gemini-3-8-flash-platform-docsWithheld with the release it exists to cite, for the same reason as its sibling source.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md

What this run does not prove

The home route is effectively full. Clean HEAD measures 1,100,526 raw critical bytes against a 1,105,000 ceiling, leaving 4,474 bytes; the three withheld records cost 9,747. At roughly 3,250 rendered bytes per release record this dataset can absorb at most one further release on the home route before the ceiling refuses it, whatever the panel says.
asset-budgets.json records measuredRaw for home as 1,080,555 while the tree measures 1,100,526. The drift of 19,971 is inside the 21,611 allowance, so it passes, but at 92% of it - a subsequent run adding as little as 1,640 bytes would fail the drift test rather than the ceiling, and would be told the ceiling was fine. ADR 0015 admits measuredRaw to the qualifying class and would have permitted regenerating it; this run was instructed to confine itself to web/src/data, and regenerating it would not have unblocked the tranche in any case, because what refuses the tranche is criticalMaxRaw, which ADR 0015 explicitly keeps out of class.
These same Gemini 3.8 Flash records were rejected by a panel on 2026-09-06 (docs-source-add 1-of-3, release-add 0-of-3) and accepted by a later panel on 2026-09-11, which then withheld them on the same byte ceiling. This run is a third independent panel and reached the 2026-09-11 result. rejection-reversals.json does not annotate the 2026-09-06 rejections and cannot be written by a refresh run, because it sits outside gate-scope ALLOWED_PATHS by design so that a run cannot absolve itself.
gate-reversals passes, but its check does not cover these rejections; they are 2 of the 186 of 204 it acknowledges it cannot see. A pass here is the absence of a finding, not a finding of absence.
The editorial dissent on the release summary was not acted on by rewording, because the chair does not amend text the panel has already read. The wording is carried to a follow-up rather than quietly fixed.
Only google-deepmind produced claims. The other 43 creators were swept by comparing fetched pages against the committed dataset for novel tokens; a change carrying no new token would not surface through that filter, so a nil result is weaker evidence than a positive one.

Follow-ups — proposed, not fixed

The home route has 4,474 bytes of headroom and each new release record costs about 3,250 rendered bytes, so the next refresh that finds more than one release will be refused exactly as this one was. This is now the binding constraint on the whole pipeline and it needs a page-weight decision, not another withheld tranche.
asset-budgets.json records a home measuredRaw 19,971 bytes below what the tree measures, which is 92% of the drift allowance. A refresh confined to web/src/data cannot correct it, so it will keep eating the allowance until someone regenerates it.
The Gemini 3.8 Flash records carry a 2026-09-06 panel rejection that rejection-reversals.json does not annotate, and no refresh run can add that annotation. Two later panels have since accepted the same records, so the register is out of step with the reviewed record.
The summary wording proposed for the Gemini 3.8 Flash release drew an editorial rejection for ranking language. Whoever lands that record should settle the wording before it ships rather than inheriting the text this run withheld.

2026-09-12-ebc58b

MAI-Code-1.1-Flash, MAI-Voice-2 and Granite 4.2 land; the nvidia bundle is refused

Scope requested: Every creator in organizations.json plus the long-tail profile: 44 scouted, none unswept. 155 pages attempted across 5 fetch rounds, 152 returning content.PublishedSep 12, 20269 edits posted · 5 items withheld

A full sweep of all 44 creators. Nine claims across microsoft and ibm were accepted unanimously by the three-rubric panel and applied: two Microsoft AI releases with their announcement sources, two IBM Granite 4.2 releases with their model-card sources, and a coverage note on the granite-4-2 family. The whole nvidia bundle was refused at the evidence gate: nvidia-nemotron-3-5-lightning-release reached 1 of the 3 accepts the long-tail policy requires, the provenance and editorial rubrics independently rejecting an unsourced status="current" that the model card behind it records as unknown, and the record is schema-invalid regardless because OpenMDW-1.1 carries no SPDX identifier. Its two unanimously accepted siblings cascaded out with it rather than landing a family with no releases.

PreflightRanClean tree, gh authenticated, no refresh pull request from a previous run still open. Drydock probed both forms: the bare name is refused by the execution policy and drydock.cmd reports 0.1.0, so the CLI is installed and blocked rather than absent.
ScoutRan155 pages attempted, 152 ok, 3 failed. Every quote byte-verbatim from a fetched page, every page sha256-hashed on the day. No search snippet was used as evidence. Three bundles carried claims; 41 creators were scouted and yielded nothing.
ReviewRanThree sub-agents, one per rubric, run in parallel and blind to each other and to the scout reasoning; each saw only the claim, its evidence and the dataset slice it touches. 36 verdicts over 12 claims. Consistency accepted all 12; editorial and provenance each rejected the same nvidia release claim independently.
GatesRanBoth bundle gates ran before anything was applied, on a clean tree, because gate-source-approval anchors on the committed dataset. gate-evidence refused the nvidia bundle at exit 1. gate-dataset, npm run validate, gate-scope and gate-ledger all passed on the applied tranche.
PublishRanTwo commits on one branch and one pull request: the dataset tranche first, then this ledger entry, which cannot be written until the pull request it must name exists. Merged by GitHub under --auto --squash once web-ci was green.
DeployNot runRecorded before the merge, so the Pages deploy had not yet run when this entry was written. Confirmed separately after merge; a failed deploy is reverted by pull request rather than left to freeze the published site on its previous build.

What was found

Scouts: 1
Pages fetched and hashed: 155
Claims proposed: 12

Claims per creator bundle, with the review threshold its profile set
Creator	Policy	Threshold	Claims
openai	pilot	2-of-3	0
anthropic	pilot	2-of-3	0
google-deepmind	pilot	2-of-3	0
meta	pilot	2-of-3	0
xai	long-tail	3-of-3	0
mistral-ai	long-tail	3-of-3	0
deepseek	long-tail	3-of-3	0
alibaba-cloud	pilot	2-of-3	0
microsoft	pilot	2-of-3	4
amazon	pilot	2-of-3	0
cohere	long-tail	3-of-3	0
ai2	long-tail	3-of-3	0
tii	long-tail	3-of-3	0
nvidia	long-tail	3-of-3	3
ai21-labs	long-tail	3-of-3	0
zhipu-ai	long-tail	3-of-3	0
moonshot-ai	long-tail	3-of-3	0
eleutherai	long-tail	3-of-3	0
lg-ai-research	long-tail	3-of-3	0
snowflake	long-tail	3-of-3	0
upstage	long-tail	3-of-3	0
ibm	long-tail	3-of-3	5
baidu	long-tail	3-of-3	0
tencent	long-tail	3-of-3	0
bytedance-seed	long-tail	3-of-3	0
stability-ai	long-tail	3-of-3	0
databricks	long-tail	3-of-3	0
minimax	long-tail	3-of-3	0
apple	long-tail	3-of-3	0
hugging-face	long-tail	3-of-3	0
01-ai	long-tail	3-of-3	0
sakana-ai	long-tail	3-of-3	0
sarvam-ai	long-tail	3-of-3	0
naver	long-tail	3-of-3	0
aleph-alpha	long-tail	3-of-3	0
reka-ai	long-tail	3-of-3	0
nous-research	long-tail	3-of-3	0
liquid-ai	long-tail	3-of-3	0
xiaomi	long-tail	3-of-3	0
ai-singapore	long-tail	3-of-3	0
kyutai	long-tail	3-of-3	0
lelapa-ai	long-tail	3-of-3	0
maritaca-ai	long-tail	3-of-3	0
openbmb	long-tail	3-of-3	0

What those claims proposed to do
Kind	Count	Effect
Add	11	Four source registrations, four releases and three nvidia records proposed as new entries; nine of the eleven were applied.
Change	1	A coverage note on the ibm-granite-4-2 family description, reflecting the two sizes added alongside it.

Not covered

Pages behind an authentication wall or a client-side rendering shell were not executed; only the served bytes were read.
Benchmark figures and vendor performance claims were not extracted at all this run, by policy: they are contested, versioned and not a fact about a release.

Degraded discovery channels

openai — official-announcement: openai.com/news/ returned HTTP 403, as did /index/gpt-4-1/ and /index/gpt-5-6/ - 3 of the run's 155 fetches. Reported rather than worked around: the run did not scrape past it or spoof a user agent. OpenAI ships no open weights, so the Hugging Face gap signal is structurally blind to it and this 403 left the run with no discovery channel for that creator at all.
zhipu-ai — official-announcement: z.ai/blog/* served a 598 to 604 byte JavaScript shell carrying no article text. Body sizes across the run range from 598 B to 1.24 MB, so "too small to be an article" discriminates here on measurement rather than on assumption.

What was evaluated

Reviewers: 3
Verdicts cast: 36
Accepted by panel: 11
Rejected by panel: 1

Deterministic gates and required checks — 7 of 8 checks passed. Exit 0 is a pass; exit 2 means the gate could not run and is never treated as one.
Check	Scope	Exit	Result
`gate-source-approval`	microsoft, ibm and nvidia bundles, pre-apply on a clean tree	0	PassEvery claim rests on an origin the committed dataset or a reviewed profile catalogue already stands behind. This run approved no new publisher.
`gate-evidence`	microsoft and ibm bundles	0	PassEvery quote resolves byte-verbatim to a fetched page whose sha256 the bundle records, and every claim meets its policy threshold.
`gate-evidence`not required	nvidia bundle	1	FailRecorded as a gate that did its job. The bundle-wide verdict refused all three nvidia claims because one of them reached 1 of 3 accepts against a unanimous long-tail threshold; nothing from the bundle was applied. Marked not-required because this failure is the mechanism that withheld the work, not a blocker on the tranche that shipped.
`gate-dataset`	applied tranche	0	PassZero failures. Referential integrity, entity boundaries and the family-has-release rule all hold across the three changed documents.
`npm run validate`	web/	0	Pass3189 of 3189 tests pass across 139 of 139 files with 0 not executed, and astro check reports 0 errors, 0 warnings, 0 hints. Exit code read from an unpiped invocation with a failing control for comparison, because a pipeline stage that stops early corrupts it.
`gate-scope`	committed HEAD and working tree	0	Passchanged=3, outOfClass empty. Only the three dataset documents moved in the data commit; this ledger entry is the one further in-class document, put in class by ADR 0006.
`gate-ledger`	this entry against the diff it describes	0	Passtranscription false, meaning the entry was written on a branch that genuinely carries the data change it reports rather than on an empty one where nothing could verify it.
`ci-preflight`	repository root, checks selected from the branch diff	0	PassSelects the pull-request checks this diff actually triggers and runs them locally. Does not cover the networked link-health sweep or the second Python interpreter, so a green preflight is not a green CI on its own.

Posted 9 edits

9 edits across 3 documents, a net change of 8 records.

Dataset documents this run changed
Document	Before	After	What changed
`web/src/data/sources.json`	291	295	Two Microsoft AI announcement posts and two IBM Granite 4.2 model cards registered as sources.
`web/src/data/releases.json`	124	128	MAI-Code-1.1-Flash, MAI-Voice-2, Granite 4.2 8B and Granite 4.2 3B, each with a publisher-stated date and so needing no dateBasis marker.
`web/src/data/families.json`	86	86	One description changed on ibm-granite-4-2 to note the sizes now covered. No family was added or removed.

Each document links to the file as this run left it, not as it stands today.

Records added

MAI-Code-1.1-Flashpassportreleasesmicrosoft-mai-code-1-1-flashDated 2026-08-11 from the announcement page, which states the date beside the headline and republishes it as JSON-LD datePublished.
MAI-Voice-2passportreleasesmicrosoft-mai-voice-2Registered from Microsoft AI's own announcement post, unanimous 3/3 against a 2-of-3 pilot threshold.
Granite 4.2 8Bpassportreleasesibm-granite-4-2-8bFrom the IBM model card, unanimous 3/3 against the unanimous long-tail threshold.
Granite 4.2 3Bpassportreleasesibm-granite-4-2-3bFrom the IBM model card, unanimous 3/3 against the unanimous long-tail threshold.

Not posted 5 items

Rejected by the review panel

nvidia-nemotron-3-5-lightning-releaseReached 1 of the 3 accepts a long-tail creator requires. The provenance and editorial rubrics rejected it independently and converged on the same reason: status="current" is stated by no quote, and the bare model card it rests on is recorded by schema.ts as status unknown. It is also schema-invalid on its own terms, failing validation on releases.128.license.spdxId because OpenMDW-1.1 carries no SPDX identifier. No reviewer was re-run and no vote was overruled.Blocked by rubric:provenancerubric:editorialgate-evidenceweb/src/data/schema.ts

Accepted by the panel, then dropped

nvidia-nemotron-3-5-lightning-sourceUnanimous 3/3, and still not applied. Landing a source without the release it pairs with fails check-bundle-pairing, and the bundle-wide evidence verdict refused the bundle whole.Blocked by nvidia-nemotron-3-5-lightning-releasecheck-bundle-pairing
nvidia-nemotron-3-5-familyUnanimous 3/3, and still not applied. A family with no releases throws at validate.ts:599 and fails the family-has-release gate in gate-dataset, so accepting it without the refused release was never available.Blocked by nvidia-nemotron-3-5-lightning-releasegate-dataset:family-has-release

Verification date deliberately held back

platform-dated-release-candidatesGLM-5.3, GLM-5.3-Flash, LFM2.5, MiniMax-Music3, Hy4-preview, Qwen-Drive, VibeVoice-ASR and Gemini 3.8 Flash were found, and each has only a hosting-platform observation for its date rather than a publisher-stated one. date-basis-policy.test.ts requires such a record to carry dateBasis and then pins the complete set of dateBasis-bearing ids with toEqual, so adding any new one fails it, and the only repair is editing a test file. That is outside the qualifying class, so the candidates were withheld rather than the class widened. The four releases that did land are unaffected because each carries a publisher-stated date and needs no marker.Blocked by web/src/data/date-basis-policy.test.tsdocs/adr/0003-an-agent-gated-data-refresh-may-auto-merge.md

Blocked by policy before it could run

bundle:nvidiaThe bundle-wide gate verdict is what the applier honours, so a single sub-threshold claim withheld all three. Recorded as policy rather than as a defect: the alternative is a panel majority on two claims overriding a deterministic gate on a third.Blocked by gate-evidence

What this run does not prove

The gates catch malformed, impossible, unreferenced and boundary-violating data. They do not catch a claim that is well-formed and simply wrong, and neither does web-ci: a plausible wrong date passes both.
All three reviewers ran claude-opus-4.8, so they share a failure mode. Three instances of one model reading one page can agree and be wrong together, and a source that is itself wrong can carry all three. The independent convergence of provenance and editorial on the nvidia rejection is evidence the panel can discriminate, not proof it always will.
A 3189-test pass is a statement about the dataset being well-formed and internally consistent, not about it being true.
Coverage of 44 of 44 creators means every creator was scouted, not that every release each creator shipped was found. A creator publishing only through a channel this run could not read - as OpenAI did, behind a 403 - can have shipped something this run reports nothing about.
The asset-budget figures were left untouched because a real production build with this tranche applied passed all 44 assertions at 1.848% drift against a 2% guard. That is a measurement of this build on this machine, not a guarantee about the deployed one.
This entry was written before the Pages deploy ran, so its deploy stage records not-run rather than a result. The deploy was confirmed separately after merge.

Follow-ups — proposed, not fixed

OpenAI's official-announcement channel has now returned 403 across this run; a creator whose only discovery channel is dark cannot be swept meaningfully however green the run looks.
The date-basis-policy test pins its expected set with toEqual, which makes every platform-dated candidate a stop for an agent run rather than a claim it can weigh. That coupling is worth a human decision, not a refresh's.

2026-09-11-cd4874

Full 44-creator sweep: 190 re-verifications published, every new record withheld at the home-route ceiling

Scope requested: All 44 creators scouted, so found.unswept is empty: nothing was passed off as unchanged that was merely not looked at. 7 creators carry a reviewed profile and ran at the 2-of-3 pilot threshold; the other 37 ran at the unanimous 3-of-3 long-tail threshold. Only web/src/data/sources.json changed, re-dating lastCheckedDate on 189 sources and correcting one notes field. No record was added, removed or re-identified, and the record count is unchanged at 291.PublishedSep 11, 2026190 edits posted · 82 items withheld

A full sweep of all 44 creators, 477 pages fetched and hashed on 2026-09-11, 398 claims proposed and 1194 verdicts cast by blind three-rubric panels - exactly three per claim. The panel accepted 357 and rejected 41. gate-evidence then refused 19 bundles whole, all of them long-tail, costing 145 individually-accepted claims. Of what survived, the run applied 190 claims and withheld all 14 accepted new records: adding them puts the home route at 1106468 bytes against a criticalMaxRaw of 1105000, and raising a ceiling is the one move ADR 0015 puts outside the qualifying class. The tranche was therefore narrowed by a stated structural rule - re-verification only, no new records - rather than by choosing which new model to hide, which would have been an unrecorded editorial judgement. web/asset-budgets.json was not touched at all. A separate evidence-reproducibility checker, two-sided controlled, found 25 of 491 evidence items unreproducible across 18 claims; all 18 had already been rejected by the blind panel, so two independent instruments converged completely.

PreflightRanClean worktree, gh authenticated, no competing refresh pull request open. Anchor d395d2c8fda531c8b633885d6fd56594b9788755 resolved and confirmed against a live git ls-remote read, then re-read at report time and found unchanged. A full baseline was taken before any edit - all gates at exit 0 and npm run validate at exit 0 with 139 files and 3189 tests - so any later failure is attributable to this run rather than inherited.
ScoutRan477 pages fetched and hashed over the bytes received on 2026-09-11 across 7 parallel scouts covering all 44 creators. 26 channels degraded and are named per creator in found.degradedChannels rather than folded into a flat error list. gate-source-approval passed 44 of 44 bundles, so no claim rests on an unapproved origin.
ReviewRan18 blind reviewer agents cast 1194 verdicts, exactly 398 x 3, so every claim received three independent reviewers and none was skipped. Each reviewer saw only the claim, its evidence, the relevant dataset slice, the creator profile and its own rubric - never another reviewer verdict, the running tally, or the scout reasoning. The chair cast no vote, broke no tie and re-ran no reviewer. 357 accepted, 41 rejected, 8 accepted over a recorded dissent.
GatesRangate-source-approval 44/44, gate-evidence 25 pass and 19 refuse, gate-dataset exit 0 over 634 records, gate-scope exit 0 with outOfClass empty and one document changed. npm run validate exit 0 with 139/139 files and 3189/3189 tests and 0 Astro errors; ci-preflight.mjs exit 0 with web-ci, skills-ci and source-link-health-tests all passing. Exit codes were read from unpiped invocations after a piped read of validate reported 1 where the true status was 0 - the Select-Object early-termination hazard, not a test failure.
PublishRanPublished as pull request #1150 with auto-merge enabled on a squash, gated by web-ci. The full per-claim evidence table and all 41 verbatim reviewer rationales are posted as comments on that pull request, because the body limit will not hold them and truncating a reviewer reasoning would defeat the point of publishing it. No bypass was used: no --force, no --admin, no skipped gate, no lowered threshold and no direct push to main.
DeployNot runNot yet observed at the moment this entry was written, because the entry ships in the same pull request it describes and the Pages deploy runs on push to main afterwards. Recorded as not-run rather than as a pass: a deploy nobody has looked at is not a deploy that succeeded.

What was found

Scouts: 7
Pages fetched and hashed: 477
Claims proposed: 398

Claims per creator bundle, with the review threshold its profile set
Creator	Policy	Threshold	Claims
01-ai	long-tail	3-of-3	5
ai-singapore	long-tail	3-of-3	20
ai2	long-tail	3-of-3	4
ai21-labs	long-tail	3-of-3	12
aleph-alpha	long-tail	3-of-3	3
alibaba-cloud	pilot	2-of-3	17
amazon	pilot	2-of-3	6
anthropic	pilot	2-of-3	19
apple	long-tail	3-of-3	7
baidu	long-tail	3-of-3	4
bytedance-seed	long-tail	3-of-3	2
cohere	long-tail	3-of-3	17
databricks	long-tail	3-of-3	2
deepseek	long-tail	3-of-3	21
eleutherai	long-tail	3-of-3	5
google-deepmind	pilot	2-of-3	22
hugging-face	long-tail	3-of-3	26
ibm	long-tail	3-of-3	5
kyutai	long-tail	3-of-3	4
lelapa-ai	long-tail	3-of-3	5
lg-ai-research	long-tail	3-of-3	5
liquid-ai	long-tail	3-of-3	3
maritaca-ai	long-tail	3-of-3	5
meta	pilot	2-of-3	25
microsoft	pilot	2-of-3	17
minimax	long-tail	3-of-3	7
mistral-ai	long-tail	3-of-3	18
moonshot-ai	long-tail	3-of-3	8
naver	long-tail	3-of-3	5
nous-research	long-tail	3-of-3	3
nvidia	long-tail	3-of-3	7
openai	pilot	2-of-3	27
openbmb	long-tail	3-of-3	12
reka-ai	long-tail	3-of-3	3
sakana-ai	long-tail	3-of-3	5
sarvam-ai	long-tail	3-of-3	4
snowflake	long-tail	3-of-3	1
stability-ai	long-tail	3-of-3	5
tencent	long-tail	3-of-3	8
tii	long-tail	3-of-3	4
upstage	long-tail	3-of-3	1
xai	long-tail	3-of-3	7
xiaomi	long-tail	3-of-3	3
zhipu-ai	long-tail	3-of-3	9

What those claims proposed to do
Kind	Count	Effect
Change	312	Re-verification of facts already in the dataset, overwhelmingly lastCheckedDate on existing sources. 190 of these were applied; the rest were lost to bundle-level gate-evidence refusal or rejected by the panel.
Add	85	Proposed new sources, families and releases. None reached the dataset: 14 accepted adds were withheld at the home-route asset ceiling, 7 were dropped on pinned test registries or their cascade, and the remainder were rejected or lost to bundle refusal.
Conflict	1	One recorded conflict between primary sources, published as a finding rather than resolved by the pipeline.

Not covered

The 9 OpenAI announcement and news pages that returned HTTP 403, and the xAI news channel that failed outright. Those creators were swept through their remaining reachable channels only, so their coverage this run is partial and is reported as partial.
No workaround was attempted against any channel that refused. A page that returns 403 is recorded as degraded rather than fetched by another route.
The 19 bundles gate-evidence refused were scouted and reviewed in full; what they did not get is an applied change. They are not unswept, and they are named individually in withheld.

Degraded discovery channels

ai-singapore — ai-singapore-gemma-sea-lion-v3-9b-docs-lowercase-404: https://github.com/aisingapore/sealion/blob/main/models/sea-lion-v3/gemma-sea-lion-v3-9b.md returned error
ai-singapore — ai-singapore-llama-sea-lion-v3-70b-docs-lowercase-404: https://github.com/aisingapore/sealion/blob/main/models/sea-lion-v3/llama-sea-lion-v3-70b.md returned error
cohere — cohere-transcribe-c4ai-model-card-probe: https://huggingface.co/CohereLabs/c4ai-transcribe-4b-v1 returned http-401
cohere — hugging-face-cohere-transcribe-c4ai-hub-record-probe: https://huggingface.co/api/models/CohereLabs/c4ai-transcribe-4b-v1 returned http-401
google-deepmind — google-gemini-3-8-flash-cyber-docs: https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash-cyber returned http-404
google-deepmind — google-gemini-3-8-flash-cyber-platform-docs: https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-8-flash-cyber returned http-404
ibm — ibm-granite-4-0-h-small-model-card: https://huggingface.co/ibm-granite/granite-4.0-h-small returned http-429
ibm — ibm-granite-4-0-h-tiny-model-card: https://huggingface.co/ibm-granite/granite-4.0-h-tiny returned http-429
ibm — ibm-granite-4-2-30b-model-card: https://huggingface.co/ibm-granite/granite-4.2-30b returned http-429
ibm — ibm-granite-4-2-technical-blog: https://huggingface.co/blog/ibm-granite/granite-4-2 returned http-429
lg-ai-research — lg-ai-research-exaone-3-5-model-card: https://huggingface.co/LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct returned http-429
mistral-ai — mistral-medium-3-5-hf-model-card-probe: https://huggingface.co/mistralai/Mistral-Medium-3.5-128B-Instruct-2604 returned http-401
openai — openai-chatgpt-images-2-5-announcement: https://openai.com/index/introducing-chatgpt-images-2-5 returned http-403
openai — openai-gpt-4-1-announcement: https://openai.com/index/gpt-4-1/ returned http-403
openai — openai-gpt-5-4-announcement: https://openai.com/index/introducing-gpt-5-4/ returned http-403
openai — openai-gpt-5-5-announcement: https://openai.com/index/introducing-gpt-5-5/ returned http-403
openai — openai-gpt-5-6-launch: https://openai.com/index/gpt-5-6/ returned http-403
openai — openai-gpt-5-6-price-update: https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/ returned http-403
openai — openai-gpt-5-6-sol-preview: https://openai.com/index/previewing-gpt-5-6-sol/ returned http-403
openai — openai-gpt-live-1-announcement: https://openai.com/index/introducing-gpt-live-1-in-the-api returned http-403
openai — openai-news: https://openai.com/news/ returned http-403
snowflake — snowflake-arctic-instruct-model-card: https://huggingface.co/Snowflake/snowflake-arctic-instruct returned http-429
tencent — tencent-hunyuanimage-3-0-instruct-distil-license: https://huggingface.co/tencent/HunyuanImage-3.0-Instruct-Distil/raw/main/LICENSE returned http-404
upstage — upstage-solar-pro-preview-model-card: https://huggingface.co/upstage/solar-pro-preview-instruct returned http-429
xai — xai-news: https://x.ai/news returned failed
zhipu-ai — zhipu-ai-glm-4-5-announcement: https://z.ai/blog/glm-4.5 returned empty-text

What was evaluated

Reviewers: 18
Verdicts cast: 1,194
Accepted by panel: 357
Rejected by panel: 41

Deterministic gates and required checks — 6 of 7 checks passed. Exit 0 is a pass; exit 2 means the gate could not run and is never treated as one.
Check	Scope	Exit	Result
`gate-source-approval`	All 44 claim bundles	0	PassEvery cited origin is in the reviewed catalogue. 44 of 44 bundles at exit 0.
`gate-evidence`	The 25 bundles that passed, which are the only bundles contributing a record to this pull request	0	PassExit 0. Every claim carries a verbatim quote, a sha256 hash and a fetch date, and every claim met the threshold its policy sets - 2-of-3 for a reviewed profile, unanimous 3-of-3 for a long-tail creator. Per ADR 0005 this gate verifies the form of a citation and never its correspondence to the cited URL, so it is not evidence that any quote is real; a separate reproducibility checker addressed that.
`gate-evidence`not required	The 19 bundles it refused, run to confirm the withholding mechanically	1	FailExit 1 on all 19, every one of them long-tail, naming the claims that fell short of the unanimous 3-of-3 bar. The run honoured every refusal exactly: nothing from any of these bundles is in this pull request, so this gate run governs no published record. It is expensive rather than wrong - refusal is per bundle, so 145 claims the panel had individually accepted were discarded alongside the blocking ones, and that cost is filed as a follow-up rather than worked around.
`gate-dataset`	The applied tranche, 634 records	0	PassSchema, referential integrity and the family-has-release rule all hold. An earlier tranche failed this gate when dropping a release orphaned its family; the family and the 4 sources whose sole referents were dropped went with it rather than landing as a data error.
`gate-scope`	Branch diff against merge-base d395d2c8fd	0	PassOne dataset document changed, web/src/data/sources.json, with outOfClass empty. web/asset-budgets.json was not touched in any field.
`npm run validate`	web/	0	Pass139 of 139 test files and 3189 of 3189 tests passed with 0 Astro errors, matching the pre-edit baseline exactly. Every asset-budget test passes.
`ci-preflight`not required	Repository root, checks selected from the branch diff	0	Passweb-ci, skills-ci and source-link-health-tests all pass locally. It does not cover the networked link-health sweep, the second Python interpreter, web-e2e or the runner itself, so it predicts CI rather than binding it.

Applied over a recorded dissent

These met their threshold and were applied. The objection stands on the record and was not overruled.

qwen3-8-flash-next-repository-last-checked-2026-09-11Editorial and entity boundariesThe current source record notes import the comparative phrase "delivers superior capabilities in coding and office tasks" into data; editorial forbids rank/winner language even when a source says it.
alibaba-qwen-drive-1-0-4b-release-addProvenanceThe attached quotes support downloadability (`The model can be downloaded from Hugging Face or ModelScope`), weights, createdAt, and Apache licensing, but none of this claim's quoted evidence states `
google-gemini-3-8-flash-release-addEditorial and entity boundariesGoogle profile says the canonical release name is the API model id and the marketing name is an alias; the proposal sets canonicalName/displayName to Gemini 3.8 Flash while treating gemini-3.8-flash a
microsoft-mai-transcribe-2-announcement-last-checked-2026-09-11Editorial and entity boundariesThe source title and claim statement import a winner/rank phrase, "fastest, most accurate and cheapest speech recognition model in the world"; editorial forbids ranking language in data or phrasing.
microsoft-mai-image-2-6-announcement-last-checked-2026-09-11Editorial and entity boundariesThe source title and claim statement import the composite ranking phrase "quality-cost frontier"; editorial forbids frontier/rank language in data or phrasing.
microsoft-mai-code-1-1-flash-announcement-source-addEditorial and entity boundariesThe proposed source title imports comparative ranking, "Better, faster, at a quarter of the cost", and the notes/quote carry quality-efficiency-cost comparisons; editorial forbids rank/composite winne
openai-gpt-image-2-5-sunburst-release-addEditorial and entity boundariesOpenAI release naming rule is dated snapshot as canonical release and dateless name as API alias; the proposal sets canonicalName/displayName to GPT-Image-2.5 Sunburst while placing gpt-image-2.5-sunb
openai-gpt-image-2-5-flare-release-addEditorial and entity boundariesOpenAI release naming rule is dated snapshot as canonical release and dateless name as API alias; the proposal sets canonicalName/displayName to GPT-Image-2.5 Flare while placing gpt-image-2.5-flare-2

Posted 190 edits

190 edits across 1 document, a net change of 0 records.

Dataset documents this run changed
Document	Before	After	What changed
`web/src/data/sources.json`	291	291	Re-dates lastCheckedDate on 189 sources across 25 creators against freshly fetched and hashed primary pages, and corrects one notes field. No record added, removed or re-identified, which is why posted.records is empty: nothing landed in releases, families or organizations this run.

Each document links to the file as this run left it, not as it stands today.

Not posted 82 items

Rejected by the review panel

ai-singapore:osi-license-index-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. Reject: the cited contentHash sha256:94c72ae924ee69fa5e1fd3481013d287afdd941a8b84bdcff8e629e279923432 does not match any stored .raw page for this run; the only osi-license-index.raw present hashes to sha256:9304b108a09fbd97c7c25474726be0cbd5760dd72be2ff9ed5b1b9b615af9965, so theBlocked by provenance
ai-singapore:ai-singapore-sea-lion-v3-description-record-three-models1/3 accepts against a threshold of 3. Dissenting rubric(s): provenance, editorial. Reject: the only attached quote says SEA-LION v3 is "a collection of 3 models (and their variants)" but does not name Gemma 9B or Llama 70B, so it does not support the proposed description's specific named list.Blocked by provenanceeditorial
ai-singapore:ai-singapore-gemma-sea-lion-v3-9b-release-add1/3 accepts against a threshold of 3. Dissenting rubric(s): provenance, editorial. Reject: the quotes support the name, 9B size, 8192 context, download availability, and "License: Gemma Community License", but no attached quote states lifecycle status "current", and the generic OSI-process quote does not state that Gemma Community License is not OSI-approved.Blocked by provenanceeditorial
ai-singapore:ai-singapore-llama-sea-lion-v3-70b-release-add1/3 accepts against a threshold of 3. Dissenting rubric(s): provenance, editorial. Reject: the attached Llama quote states "70 billion parameters" and download availability, but no quote states the proposed 128000-token context window or lifecycle status "current", and the generic OSI-process quote does not state that the Llama 3.1 Community License is not OSI-Blocked by provenanceeditorial
ai-singapore:ai-singapore-llama-sea-lion-v3-8b-sibling-ids-add-v3-peers2/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. Reject: the quote says only that SEA-LION v3 is "a collection of 3 models (and their variants)"; it does not name the proposed sibling ids Gemma-SEA-LION-v3-9B and Llama-SEA-LION-v3-70B.Blocked by provenance
ai-singapore:ai-singapore-sea-lion-v4-5-family-add1/3 accepts against a threshold of 3. Dissenting rubric(s): provenance, editorial. Reject: the quotes support v4.5 as "foundational, agentic, and multimodal" and identify Qwen/Gemma sources, but none of the attached quotes states the proposed "coding" category.Blocked by provenanceeditorial
ai21-labs:osi-approved-licenses-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The cited contentHash sha256:b4c6f6a756d1eba09084fd3db309aec635290e58fb4ddda677ed92a5e3837343 does not match the fetched osi-approved-licenses.raw bytes, which hash to sha256:087690e000fec800e0e5e6394a74f0391e65b60e409df4c4b8d55b65fcb571ff.Blocked by provenance
ai21-labs:ai21-labs-jamba-instruct-release-add2/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The quote says "Jamba-Instruct ... is in public preview on the AI21 Platform", so the proposed status unknown is not supported; the source states a preview lifecycle value.Blocked by provenance
apple:apple-openelm-license-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The quoted string "Disclaimer: IMPORTANT: This Apple Machine Learning Research Model is" is generic Apple model-license boilerplate; it does not name OpenELM, OpenELM-3B-Instruct, or any source-specific license title, so it does not prove the catalogued OpenELM license source stiBlocked by provenance
apple:apple-fastvlm-license-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The quoted string "Disclaimer: IMPORTANT: This Apple Machine Learning Research Model is" is generic Apple model-license boilerplate; it does not name FastVLM, FastVLM-7B, or any source-specific license title, so it does not prove the catalogued FastVLM license source still identiBlocked by provenance
cohere:cohere-command-a-plus-announcement-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The quote is only the generic banner 'Skip to content AI for Empowerment: Your freedom. Your focus...' and does not identify the Command A+ announcement page, so it does not re-verify that catalogued source.Blocked by provenance
cohere:cohere-transcribe-announcement-source2/3 accepts against a threshold of 3. Dissenting rubric(s): editorial. Reject: the proposed source title and claim statement import "state-of-the-art" into ModelTree source data/phrasing; the editorial rubric forbids rank or winner language even when it appears in an announcement title.Blocked by editorial
cohere:cohere-north-small-translate-announcement-source2/3 accepts against a threshold of 3. Dissenting rubric(s): editorial. Reject: the proposed source title and claim statement import the rank term "leading" into ModelTree source data/phrasing; the editorial rubric forbids rank or winner language.Blocked by editorial
cohere:cohere-north-small-translate-model-card-source2/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The quote is tokenizer/chat-template text beginning with eos_token/pad_token/platform_instruction_override; it does not name North Small Translate or CohereLabs/North-Small-Translate-1.0, so it does not support adding that model-card source.Blocked by provenance
cohere:cohere-transcribe-03-2026-release2/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The quotes support ASR, audio-in/text-out, 2B parameters, and download availability, but no model quote states Apache 2.0 as the model licence or hosted/API/Model Vault access for accessType 'both'; the OSI page only proves Apache approval.Blocked by provenance
cohere:cohere-north-small-translate-1-0-release2/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The model-card quote supports an open-weights research MoE with 25B active and 218B total parameters for translation, but no provided quote states the CC BY-NC 4.0 licence or the proposed 16K input/output context windows.Blocked by provenance
cohere:cohere-parse-v5-0-release1/3 accepts against a threshold of 3. Dissenting rubric(s): provenance, editorial. The quote supports Cohere Parse as a document-processing vision language model, but it does not state the proposed parse-v5.0 API alias/version, current hosted availability, or proprietary-hosted access type.Blocked by provenanceeditorial
deepseek:deepseek-v4-1-family1/3 accepts against a threshold of 3. Dissenting rubric(s): provenance, editorial. The quoted source names the release 'DeepSeek-V4.1-Flash'; it does not state a separate 'DeepSeek-V4.1' family, and the hub-record quote provided does not include createdAt to support the proposed 2026-09-10 firstReleaseDate.Blocked by provenanceeditorial
deepseek:deepseek-v4-1-flash-release1/3 accepts against a threshold of 3. Dissenting rubric(s): provenance, editorial. The model-card quote supports multimodal MoE, 552B backbone parameters, image/text input, text output, and one-million-token context, but the provided hub-record quote does not include createdAt, so the proposed 2026-09-10 releaseDate/dateBasis is not quoted.Blocked by provenanceeditorial
deepseek:deepseek-v4-pro-0813-release1/3 accepts against a threshold of 3. Dissenting rubric(s): provenance, editorial. The model-card quote says DeepSeek-V4-Pro-0813 is the official release superseding the preview, but neither quote states the proposed 2026-08-13 repository-created release date; the hub-record quote shown does not include createdAt.Blocked by provenanceeditorial
deepseek:deepseek-v4-flash-0731-release1/3 accepts against a threshold of 3. Dissenting rubric(s): provenance, editorial. The model-card quote says DeepSeek-V4-Flash-0731 is the official release superseding the preview, but neither quote states the proposed 2026-07-31 repository-created release date; the hub-record quote shown does not include createdAt.Blocked by provenanceeditorial
hugging-face:hugging-face-gguf-docs-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The quoted string begins with the documentation navigation text "?? View all docs AWS Trainium & Inferentia..." and does not name GGUF or the GGUF docs page; it is boilerplate navigation, not evidence that the catalogued GGUF source still identifies itself.Blocked by provenance
ibm:ibm-granite-4-0-h-micro-release-add2/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. For open-weight access the evidence has "open sourced under a standard Apache 2.0 license" and "Download language models", but no quote states weights are downloadable; the OSI Apache evidence hash also does not match the fetched raw bytes.Blocked by provenance
ibm:ibm-granite-4-0-micro-release-add2/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. For open-weight access the evidence has "open sourced under a standard Apache 2.0 license" and "Download language models", but no quote states weights are downloadable; the OSI Apache evidence hash also does not match the fetched raw bytes.Blocked by provenance
kyutai:osi-license-index-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. Reject: the cited contentHash sha256:e68145cd8074237ae72422bf5eb1d6559f1c30896ab494cb82265edce4eb89f7 does not match any stored .raw page for this run; the only osi-license-index.raw present hashes to sha256:9304b108a09fbd97c7c25474726be0cbd5760dd72be2ff9ed5b1b9b615af9965, so theBlocked by provenance
lelapa-ai:osi-license-index-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. Reject: the cited contentHash sha256:f0205cc1dd07ecbca01b3343be9bbaa9957840aeb864a753a43c24718bb8d33f does not match any stored .raw page for this run; the only osi-license-index.raw present hashes to sha256:9304b108a09fbd97c7c25474726be0cbd5760dd72be2ff9ed5b1b9b615af9965, so theBlocked by provenance
maritaca-ai:osi-license-index-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. Reject: the cited contentHash sha256:a2c93600a45439ef11e3d486926e007d79eb5862dde2fb6305113a5ea08dd4e7 does not match any stored .raw page for this run; the only osi-license-index.raw present hashes to sha256:9304b108a09fbd97c7c25474726be0cbd5760dd72be2ff9ed5b1b9b615af9965, so theBlocked by provenance
minimax:minimax-model-license-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The quoted string "Model Release Date: 15 January 2025" is present, but it does not name MiniMax, the MiniMax Model License, or the catalogued license source; the quote alone supports only a date line, not re-verification of that source.Blocked by provenance
mistral-ai:mistral-medium-3-5-128b-release1/3 accepts against a threshold of 3. Dissenting rubric(s): provenance, editorial. The quotes support the name, 128B size, open weights, and Modified MIT wording, but no provided quote states the proposed 256k context window, image input, public-preview lifecycle, or hosted/API access required by accessType 'both'.Blocked by provenanceeditorial
mistral-ai:mistral-voxtral-tts-4b-2603-release1/3 accepts against a threshold of 3. Dissenting rubric(s): provenance, editorial. The quotes state Voxtral TTS, 4B parameters, text-to-speech/voice-cloning details, open weights/BF16 weights, and CC BY-NC wording, but none states the proposed hosted API/Studio access needed for accessType 'both' or the exact release date.Blocked by provenanceeditorial
moonshot-ai:osi-license-index-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The cited contentHash sha256:66f0bc037700f668da07e931e1a703c2347a8d32b396e2eeb36b95402aac0b71 does not match the fetched osi-license-index.raw bytes, which hash to sha256:9304b108a09fbd97c7c25474726be0cbd5760dd72be2ff9ed5b1b9b615af9965.Blocked by provenance
moonshot-ai:moonshot-ai-kimi-k2-instruct-0905-model-card-source-add2/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The cited contentHash sha256:12d67666d2a8b6790e738ab87925848d97b4628350bfdf88452106c734092278 does not match moonshot-ai-kimi-k2-instruct-0905-model-card.raw sha256:12d676ae66ec2b3dab35b76660626e94121b85b8868d98bf1df047b78b9f7ec4.Blocked by provenance
moonshot-ai:hugging-face-kimi-k2-instruct-0905-hub-record-source-add2/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The cited contentHash sha256:35b6f803c9e89f1175ad86f43b2779fed6468d9f40e5f749b8120bd4fb7e3f3e does not match hugging-face-kimi-k2-instruct-0905-hub-record.raw sha256:35b6f8eb5a864d7f8aca36dbe2cdfc16b4c69184e572a13105a1e3f6adacdd6f.Blocked by provenance
moonshot-ai:moonshot-ai-kimi-k2-instruct-0905-release-add2/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The load-bearing 0905 model-card, Hub-record, and OSI-index evidence hashes do not match the fetched raw files, so the quoted strings are not tied to the cited bytes for the release, date, access, or OSI-false fields.Blocked by provenance
naver:naver-hyperclova-x-seed-announcement-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The quoted sentence "Businesses and research institutions... can download these models" is anaphoric and does not name HyperCLOVA X SEED, NAVER, or the announcement title; read alone it does not identify the catalogued source.Blocked by provenance
sakana-ai:sakana-ai-evollm-jp-model-card-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The quoted sentence "This model is provided for research and development purposes only..." is a generic model-card disclaimer and does not name EvoLLM-JP-v1-7B or SakanaAI, so it does not prove the catalogued model-card source still identifies itself.Blocked by provenance
sakana-ai:sakana-ai-evollm-jp-license-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The quoted sentence "These license terms are an agreement between you and Microsoft Corporation..." is generic Microsoft license boilerplate and does not name EvoLLM-JP-v1-7B, SakanaAI, or the catalogued license source.Blocked by provenance
sarvam-ai:sarvam-ai-sarvam-m-blog-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The quoted sentence "Download the model from Hugging Face, try it on our playground, and build with our APIs" says only "the model" and does not name Sarvam-M or the announcement source; read alone it does not identify the catalogued source.Blocked by provenance
tencent:tencent-hunyuanimage-3-0-instruct-release2/3 accepts against a threshold of 3. Dissenting rubric(s): editorial. Reject: the release naming keeps `HunyuanImage-3.0-Instruct` and the Hub download separate, but the proposed `intendedUse` adds `chain-of-thought-style visual reasoning`; the attached quotes say only `Instruct (with reasoning)` / `Instruction reasoning` and image-to-image generatBlocked by editorial
upstage:osi-license-mit-last-checked-2026-09-112/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The cited contentHash sha256:64b9a6077ff220cdb35ad7f6f4fa3f91bdd95a72112de4d1d7233c8356e083bf does not match the fetched osi-license-mit.raw bytes, which hash to sha256:fccedc89f37399eefd47c15460bf41426cc3a3d296ea9ee3d5947aa055724716.Blocked by provenance
zhipu-ai:zhipu-ai-glm-4-5-release-add2/3 accepts against a threshold of 3. Dissenting rubric(s): provenance. The MIT OSI evidence is not tied to the fetched bytes: the cited sha256:64b9a6077ff220cdb35ad7f6f4fa3f91bdd95a72112de4d1d7233c8356e083bf does not match osi-license-mit.raw sha256:fccedc89f37399eefd47c15460bf41426cc3a3d296ea9ee3d5947aa055724716.Blocked by provenance

Accepted by the panel, then dropped

alibaba-qwen-drive-1-0-4b-release-addAdding this release puts a twelfth id into the deliberately pinned registry in web/src/data/date-basis-policy.test.ts (dateBasis 'platform-repository-created'), and a sixteenth into the pinned EXPECTED_GAP_IDS in web/src/data/parameter-gap-field.test.ts (the id states '4B' but the record carries no parameters block). Both lists are hard-coded in TypeScript test files and both are pinned on purpose - date-basis-policy.test.ts states 'Pinned rather than derived. Deriving the expected set from the data would make this test agree with whatever the data says, which is not a check.' Landing the releBlocked by web/src/data/parameter-gap-field.test.ts
lg-ai-research-exaone-4-0-1-2b-release-addThe id states '2B' and the record carries no parameters block, so it enters the pinned EXPECTED_GAP_IDS registry in web/src/data/parameter-gap-field.test.ts and additionally fails that file's 'the parameter-count explanation must be in summary' assertion, which requires the phrase 'parameter count' in the summary field. Satisfying either would mean editing a .ts test file or editing the claim's proposed summary to make it pass - the first is out of class under ADR 0003, the second is explicitly forbidden by claim-bundle.md ('Never edit a claim to make it pass - that is the run overruling its oBlocked by web/src/data/parameter-gap-field.test.ts
qwen-drive-family-addCascade from dropping alibaba-qwen-drive-1-0-4b-release-add, which was the family's only release. gate-dataset refuses a family with no releases outright ([family-has-release] 'no release belongs to this family, so the site cannot be built'), and the dataset cannot express 'announced but unreleased'. Dropped together with the release rather than left as a data error.Blocked by web/src/data/parameter-gap-field.test.ts
qwen-drive-1-0-4b-model-card-source-addCascade. This source was added solely to support the dropped Qwen-Drive release and family, and nothing else cites it. Landing a source no record references adds an uncited row to the catalogue rather than evidence.Blocked by web/src/data/parameter-gap-field.test.ts
hugging-face-qwen-drive-1-0-4b-hub-record-source-addCascade. Added solely to support the dropped Qwen-Drive release and family; nothing else cites it.Blocked by web/src/data/parameter-gap-field.test.ts
qwen-drive-1-0-repository-source-addCascade. Added solely to support the dropped Qwen-Drive release and family; nothing else cites it.Blocked by web/src/data/parameter-gap-field.test.ts
lgai-exaone-4-0-1-2b-model-card-source-addCascade. Added solely to support the dropped EXAONE 4.0.1 2B release; nothing else cites it.Blocked by web/src/data/parameter-gap-field.test.ts

Sources conflict, so no value changed

ai-singapore:ai-singapore-qwen-sea-lion-v4-5-27b-it-license-conflictRecorded as a conflict rather than resolved. The fetched Qwen-SEA-LION-v4.5-27B-IT documentation and Hub record disagree on the license value.Blocked by conflicting primary sources

Blocked by policy before it could run

bundle:hugging-facegate-evidence refused this long-tail bundle whole on 1 sub-threshold claim(s), discarding 25 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidencehugging-face-gguf-docs-last-checked-2026-09-11 (2/3 against 3)
bundle:deepseekgate-evidence refused this long-tail bundle whole on 4 sub-threshold claim(s), discarding 17 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidencedeepseek-v4-1-family (1/3 against 3)deepseek-v4-1-flash-release (1/3 against 3)deepseek-v4-pro-0813-release (1/3 against 3)deepseek-v4-flash-0731-release (1/3 against 3)
bundle:mistral-aigate-evidence refused this long-tail bundle whole on 2 sub-threshold claim(s), discarding 16 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidencemistral-medium-3-5-128b-release (1/3 against 3)mistral-voxtral-tts-4b-2603-release (1/3 against 3)
bundle:ai-singaporegate-evidence refused this long-tail bundle whole on 6 sub-threshold claim(s), discarding 13 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidenceosi-license-index-last-checked-2026-09-11 (2/3 against 3)ai-singapore-sea-lion-v3-description-record-three-models (1/3 against 3)ai-singapore-gemma-sea-lion-v3-9b-release-add (1/3 against 3)ai-singapore-llama-sea-lion-v3-70b-release-add (1/3 against 3)
bundle:ai21-labsgate-evidence refused this long-tail bundle whole on 2 sub-threshold claim(s), discarding 10 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidenceosi-approved-licenses-last-checked-2026-09-11 (2/3 against 3)ai21-labs-jamba-instruct-release-add (2/3 against 3)
bundle:coheregate-evidence refused this long-tail bundle whole on 7 sub-threshold claim(s), discarding 10 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidencecohere-command-a-plus-announcement-last-checked-2026-09-11 (2/3 against 3)cohere-transcribe-announcement-source (2/3 against 3)cohere-north-small-translate-announcement-source (2/3 against 3)cohere-north-small-translate-model-card-source (2/3 against 3)
bundle:zhipu-aigate-evidence refused this long-tail bundle whole on 1 sub-threshold claim(s), discarding 8 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidencezhipu-ai-glm-4-5-release-add (2/3 against 3)
bundle:tencentgate-evidence refused this long-tail bundle whole on 1 sub-threshold claim(s), discarding 7 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidencetencent-hunyuanimage-3-0-instruct-release (2/3 against 3)
bundle:minimaxgate-evidence refused this long-tail bundle whole on 1 sub-threshold claim(s), discarding 6 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidenceminimax-model-license-last-checked-2026-09-11 (2/3 against 3)
bundle:applegate-evidence refused this long-tail bundle whole on 2 sub-threshold claim(s), discarding 5 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidenceapple-openelm-license-last-checked-2026-09-11 (2/3 against 3)apple-fastvlm-license-last-checked-2026-09-11 (2/3 against 3)
bundle:lelapa-aigate-evidence refused this long-tail bundle whole on 1 sub-threshold claim(s), discarding 4 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidenceosi-license-index-last-checked-2026-09-11 (2/3 against 3)
bundle:maritaca-aigate-evidence refused this long-tail bundle whole on 1 sub-threshold claim(s), discarding 4 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidenceosi-license-index-last-checked-2026-09-11 (2/3 against 3)
bundle:moonshot-aigate-evidence refused this long-tail bundle whole on 4 sub-threshold claim(s), discarding 4 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidenceosi-license-index-last-checked-2026-09-11 (2/3 against 3)moonshot-ai-kimi-k2-instruct-0905-model-card-source-add (2/3 against 3)hugging-face-kimi-k2-instruct-0905-hub-record-source-add (2/3 against 3)moonshot-ai-kimi-k2-instruct-0905-release-add (2/3 against 3)
bundle:navergate-evidence refused this long-tail bundle whole on 1 sub-threshold claim(s), discarding 4 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidencenaver-hyperclova-x-seed-announcement-last-checked-2026-09-11 (2/3 against 3)
bundle:ibmgate-evidence refused this long-tail bundle whole on 2 sub-threshold claim(s), discarding 3 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidenceibm-granite-4-0-h-micro-release-add (2/3 against 3)ibm-granite-4-0-micro-release-add (2/3 against 3)
bundle:kyutaigate-evidence refused this long-tail bundle whole on 1 sub-threshold claim(s), discarding 3 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidenceosi-license-index-last-checked-2026-09-11 (2/3 against 3)
bundle:sakana-aigate-evidence refused this long-tail bundle whole on 2 sub-threshold claim(s), discarding 3 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidencesakana-ai-evollm-jp-model-card-last-checked-2026-09-11 (2/3 against 3)sakana-ai-evollm-jp-license-last-checked-2026-09-11 (2/3 against 3)
bundle:sarvam-aigate-evidence refused this long-tail bundle whole on 1 sub-threshold claim(s), discarding 3 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidencesarvam-ai-sarvam-m-blog-last-checked-2026-09-11 (2/3 against 3)
bundle:upstagegate-evidence refused this long-tail bundle whole on 1 sub-threshold claim(s), discarding 0 claim(s) the panel had accepted. The gate refuses per bundle, not per claim; the run obeyed it rather than applying a subset it had not been authorised to pick.Blocked by gate-evidenceosi-license-mit-last-checked-2026-09-11 (2/3 against 3)
google-gemini-3-8-flash-docsAccepted by the panel and not applied. Adding the new records takes the home route's critical payload to 1106468 bytes against criticalMaxRaw 1105000, over by 1468. Raising that ceiling is out of the qualifying class under ADR 0015, so the run withheld every new record rather than choosing which to hide.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md
google-gemini-3-8-flash-platform-docsAccepted by the panel and not applied. Adding the new records takes the home route's critical payload to 1106468 bytes against criticalMaxRaw 1105000, over by 1468. Raising that ceiling is out of the qualifying class under ADR 0015, so the run withheld every new record rather than choosing which to hide.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md
google-gemini-3-8-flash-cyber-announcementAccepted by the panel and not applied. Adding the new records takes the home route's critical payload to 1106468 bytes against criticalMaxRaw 1105000, over by 1468. Raising that ceiling is out of the qualifying class under ADR 0015, so the run withheld every new record rather than choosing which to hide.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md
microsoft-mai-code-1-1-flash-announcementAccepted by the panel and not applied. Adding the new records takes the home route's critical payload to 1106468 bytes against criticalMaxRaw 1105000, over by 1468. Raising that ceiling is out of the qualifying class under ADR 0015, so the run withheld every new record rather than choosing which to hide.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md
microsoft-mai-code-1-1-flash-model-pageAccepted by the panel and not applied. Adding the new records takes the home route's critical payload to 1106468 bytes against criticalMaxRaw 1105000, over by 1468. Raising that ceiling is out of the qualifying class under ADR 0015, so the run withheld every new record rather than choosing which to hide.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md
openai-gpt-live-1-docsAccepted by the panel and not applied. Adding the new records takes the home route's critical payload to 1106468 bytes against criticalMaxRaw 1105000, over by 1468. Raising that ceiling is out of the qualifying class under ADR 0015, so the run withheld every new record rather than choosing which to hide.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md
openai-gpt-image-2-5-sunburst-docsAccepted by the panel and not applied. Adding the new records takes the home route's critical payload to 1106468 bytes against criticalMaxRaw 1105000, over by 1468. Raising that ceiling is out of the qualifying class under ADR 0015, so the run withheld every new record rather than choosing which to hide.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md
openai-gpt-image-2-5-flare-docsAccepted by the panel and not applied. Adding the new records takes the home route's critical payload to 1106468 bytes against criticalMaxRaw 1105000, over by 1468. Raising that ceiling is out of the qualifying class under ADR 0015, so the run withheld every new record rather than choosing which to hide.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md
openai-gpt-liveAccepted by the panel and not applied. Adding the new records takes the home route's critical payload to 1106468 bytes against criticalMaxRaw 1105000, over by 1468. Raising that ceiling is out of the qualifying class under ADR 0015, so the run withheld every new record rather than choosing which to hide.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md
google-gemini-3-8-flashAccepted by the panel and not applied. Adding the new records takes the home route's critical payload to 1106468 bytes against criticalMaxRaw 1105000, over by 1468. Raising that ceiling is out of the qualifying class under ADR 0015, so the run withheld every new record rather than choosing which to hide.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md
microsoft-mai-code-1-1-flashAccepted by the panel and not applied. Adding the new records takes the home route's critical payload to 1106468 bytes against criticalMaxRaw 1105000, over by 1468. Raising that ceiling is out of the qualifying class under ADR 0015, so the run withheld every new record rather than choosing which to hide.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md
openai-gpt-live-1Accepted by the panel and not applied. Adding the new records takes the home route's critical payload to 1106468 bytes against criticalMaxRaw 1105000, over by 1468. Raising that ceiling is out of the qualifying class under ADR 0015, so the run withheld every new record rather than choosing which to hide.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md
openai-gpt-image-2-5-sunburstAccepted by the panel and not applied. Adding the new records takes the home route's critical payload to 1106468 bytes against criticalMaxRaw 1105000, over by 1468. Raising that ceiling is out of the qualifying class under ADR 0015, so the run withheld every new record rather than choosing which to hide.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md
openai-gpt-image-2-5-flareAccepted by the panel and not applied. Adding the new records takes the home route's critical payload to 1106468 bytes against criticalMaxRaw 1105000, over by 1468. Raising that ceiling is out of the qualifying class under ADR 0015, so the run withheld every new record rather than choosing which to hide.Blocked by web/asset-budgets.json criticalMaxRawdocs/adr/0015-asset-budget-measurements-are-in-class-their-ceilings-are-not.md

What this run does not prove

No human read any of these claims. The three-rubric panel and the deterministic gates are the whole of the oversight, and 2-of-3 buys independence of reasoning, not independence of training: three instances of the same model family reading one page share a failure mode, and a source that is itself wrong can carry all three.
A green run does not mean the dataset is complete. 145 individually-accepted claims were discarded by bundle-level gate-evidence refusal, and every one of the 19 refused bundles was long-tail, where a single dissent blocks.
The 14 withheld records are accepted, sourced and absent from the site. Readers will not see Gemini 3.8 Flash, MAI-Code-1.1-Flash or three OpenAI releases until the home-route payload problem is fixed, and nothing on the site says so.
Coverage for OpenAI and xAI is partial: 9 OpenAI pages returned 403 and the xAI news channel failed. This run is not evidence that those creators did or did not change on the channels that refused.
A separate checker found 25 of 491 evidence items unreproducible. It converged completely with the panel this run, but that convergence is a measurement of one run and not a guarantee that the scout hash-recording defect is harmless.
The deploy stage had not run when this entry was written, so nothing here reports that the site built or published. That is recorded as not-run rather than assumed.

Follow-ups — proposed, not fixed

The home route's criticalMaxRaw is exhausted: an ordinary refresh tranche no longer fits, and this run withheld 14 accepted records to stay inside it. The named remedy is to project the home page's island props rather than inlining whole records, as was done for /tree. Until then every refresh is capped at re-verification.
gate-evidence refuses at bundle granularity, so one sub-threshold claim discards every other claim in its bundle. Hugging Face lost 25 accepted claims to a single blocking one, and 145 were lost run-wide. Claim-level refusal would recover them without weakening any threshold.
The scout recorded content hashes that do not reproduce for 25 of 491 evidence items. The panel independently rejected every affected claim this run, but the defect is in the recording step and should not be left resting on that coincidence.
licence-identity.test.ts asserts that git status --porcelain src/data is empty, so it fails locally for every refresh before the commit regardless of cause, including for the ledger entry itself. It conflates "the CLI wrote to the dataset" with "the tree is dirty for any reason".
The pinned registries in date-basis-policy.test.ts and parameter-gap-field.test.ts structurally block an unattended refresh from adding any release whose id states a parameter count or whose date rests on a platform record. The pins are deliberate and correct; what is missing is a route for an agent to land such a release without editing a guard.