ModelTreeAI Model Lineage
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
23
Pages fetched
2,128
Claims proposed
1,352
Edits published
611
Items withheld
439

Showing 21–23 of 23 runs

  1. 2026-08-27-4f1c9e

    Long-tail refresh 2026-08-27 — Microsoft, Mistral AI, xAI, Cohere

    Scope requested: Four creators with no reviewed profile — Microsoft, Mistral AI, xAI and Cohere — under the long-tail unanimous 3-of-3 policy, restricted to the eight origins PR #445 approved for them.Ran, changed nothing0 edits posted · 7 items withheld

    Researched four long-tail creators to populate the Others branch of the tree and published none of them. Every fact came from the origins PR #445 approved for these creators, fetched and hashed rather than read from search results. Microsoft and xAI were withheld before review: no approved-origin page states an access type or modality set for MAI-Thinking-1, and both are required fields, while x.ai now self-describes as "SpaceXAI LLC" on both of its approved origins and no approved page names a Grok family. Mistral AI and Cohere reached a review panel across five rounds. A final round re-scouted Cohere against a model reference on an already-approved origin that carries an explicit Status column and a machine-readable specification table, which answered the earlier objections about status, categories and the creator relationship; the resulting organization and release claims were accepted by the consistency and editorial rubrics but not by provenance, so they did not reach the unanimous 3-of-3 long-tail bar. Because model-tree.ts renders a creator only when it has at least one release, the organizations and families that were accepted would have added records that render nowhere, so they were dropped rather than applied. The dataset is unchanged; this entry is the only edit.

    1. PreflightRanClean tree rebased onto main at 843f2d4. main moved three times during this run — f665967 to 3e6f3e3 to 7fd5ea2 to 843f2d4 — and every gate below was recomputed against the final anchor rather than carried forward, because a gate result binds to the commit it was computed against. drydock is not on PATH, so the manual posture applied. ADR 0003 and all four gate scripts were read directly rather than from a summary; gate-scope ALLOWED_PATHS was re-read after #448 grew the qualifying class from 9 files to 11.
    2. ScoutRan28 pages fetched over real HTTP from approved origins only and hashed with sha256; every quote was verified verbatim against those bytes before any claim was written, and a mechanical check rejected three round-4 quotes that had been copied from a stripped-text rendering rather than from the hashed HTML. x.ai/news returned a Cloudflare 403 and was never readable. docs.mistral.ai/getting-started/models/models_overview/ redirected outside that origin’s allowed_paths, so it was fetched but deliberately not cited, leaving Mistral without a context-window figure. The round-5 Cohere re-scout admitted no new host and no new page: it re-read bytes already fetched and already proposed as a source, and quoted the parts of them that state the facts under review.
    3. ReviewRanFive blind panel rounds, three rubrics each, run on three different model families with the rubric-to-family assignment rotated every round so no rubric kept one family. Reviewers never saw the scout’s reasoning, each other’s verdicts, or the running tally. No threshold was lowered and no reviewer was re-run on an unchanged claim; each corrected claim was given a new id and put to a fresh panel, and claims already accepted unanimously kept their original verdicts rather than being re-reviewed. The run stopped after round 5 rather than re-panelling again: the surviving provenance objections are about how much of a dataset record a quote must literally state, which no further evidence can resolve, and re-sampling reviewers until three agree is the failure the panel exists to prevent.
    4. GatesRangate-evidence and gate-source-approval were run before anything was applied, anchored on the committed dataset at merge-base 843f2d4 so the run could not approve its own writes, and re-run at that anchor after the round-5 bundle was written because a gate result binds to the commit it was computed against. gate-dataset, npm run validate and gate-scope were run afterwards. gate-evidence failed exactly as the panel did, naming the sub-threshold claims and nothing else.
    5. PublishRanAn ordinary pull request carrying this log entry and no dataset change. Auto-merge was not requested and not eligible: refresh-runs.json is outside the gate-scope qualifying class, so ADR 0003 does not cover this change and a human merges it.
    6. DeployNot applicableNo dataset change to deploy. The rendered tree is byte-identical to the one already on main; only the /refresh/ page gains this entry.

    What was found

    Scouts
    4
    Pages fetched and hashed
    28
    Claims proposed
    12
    Claims per creator bundle, with the review threshold its profile set
    CreatorPolicyThresholdClaims
    mistral-ailong-tail3-of-35
    coherelong-tail3-of-37
    What those claims proposed to do
    KindCountEffect
    Add12Would have added two organizations, two families, two releases, two publishers and four sources. None was applied.

    Not covered

    • Microsoft produced no bundle that reached final adjudication: MAI-Thinking-1 has no approved-origin statement of accessType or modalities, and both are required by releaseSchema.
    • xAI produced no bundle that reached final adjudication: the SpaceXAI/xAI naming conflict is unresolved across both approved origins and no approved page names a Grok family.
    • azure.microsoft.com, learn.microsoft.com and api-docs.deepseek.com were deliberately not cited. Microsoft as a serving platform and product vendor is a different entity from Microsoft as a model creator.
    • No context-window figure for any Mistral release: the only page stating one redirects outside the approved allowed_paths.
    • No approved Cohere origin states OSI approval for the Apache 2.0 licence, and licenseSchema requires osiApproved whenever a license object is present, so the whole object was omitted rather than guessed. opensource.org is not an approved origin and was not cited to close the gap.
    • No approved Cohere origin states the Command A family’s first release date in prose; the only available date is the publication timestamp of the launch announcement.

    What was evaluated

    Reviewers
    3
    Verdicts cast
    45
    Accepted by panel
    8
    Rejected by panel
    4
    Deterministic gates and required checks — 3 of 5 checks passed. Exit 0 is a pass; exit 2 means the gate could not run and is never treated as one.
    CheckScopeExitResult
    gate-evidence2 bundles, 12 claims1FailFailed on exactly the four claims the panel put below the unanimous 3-of-3 long-tail threshold: mi4-release-large-3-add, co5-organization-add, co5-family-command-a-add and co5-release-command-a-plus-add. Every quote was present, verbatim against the hashed bytes and over the 24-character minimum; the failure is the review threshold, not the evidence format. Nothing was applied, so nothing shipped past it.
    gate-source-approval2 bundles, 57 citations0PassAnchored on the committed dataset at merge-base 843f2d4 with refs/remotes/origin/main, computed by the gate rather than supplied, so the run never approved its own source. 24 already-trusted origins; every citation in both bundles, including the enlarged round-5 Cohere bundle, landed on origins the reviewed profile catalogues merged in PR #445 already stand behind. No new host was admitted. Re-run after #448 changed sources.json, after #451 moved main, and again after the round-5 bundle was written, because a gate result binds to its anchor.
    gate-datasetworking tree0PassThe dataset is unchanged by this run and remains internally coherent. No orphaned source, no dangling reference, no release predating its family.
    npm run validateweb/0PassFull test suite and Astro/TypeScript diagnostics green, including the refresh-log schema test that this new entry must satisfy.
    gate-scopebranch vs merge-base 843f2d41FailReports web/src/data/refresh-runs.json as outOfClass, which is correct and expected: the ADR 0003 qualifying class is the 11 dataset documents raw.ts composes, and the refresh log is deliberately not one of them. The gate was not modified and the file was not added to ALLOWED_PATHS; the consequence is simply that this change is outside ADR 0003 and merges the ordinary way, with a human, rather than auto-merging.

    Posted 0 edits

    Nothing reached the dataset. No branch, no commit, no pull request.

    Not posted 7 items

    Rejected by the review panel

    • mistral-large-3The only Mistral release claim, rejected by the provenance rubric in two consecutive rounds on different grounds and by different model families. Round 3 objected to an intendedUse clause asserting the model is "demanding to self-host", which no quote stated; that clause was deleted. Round 4 then objected that the quotes do not state the status, category or open-weight access type. Consistency and editorial accepted it both times. Under the unanimous 3-of-3 long-tail bar it did not carry.Blocked by mi3-release-large-3-addmi4-release-large-3-add
    • cohere-command-a-plus-05-2026The only Cohere release claim, put to three separate panels on corrected inputs and never unanimous. Round 3 caught a real schema error — contextWindow given as an object where releaseSchema requires a positive integer — which was corrected. Round 4 provenance objected that the quotes state neither osiApproved nor the dataset categories nor the intendedUse guidance. Round 5 rewrote the claim against Cohere’s own specification table, dropped the licence object entirely rather than guess osiApproved, and rebuilt categories, status, context limits and intendedUse from quoted text; consistency and editorial both accepted it, and editorial specifically credited the dated canonical name and the exclusion of Azure Foundry ids and North product metrics. Provenance still rejected, holding that mapping the table’s "Live" to the schema’s "current", normalising "128K" to 128000, and reading a release date off the announcement’s datePublished are each inferences the quotes do not contain.Blocked by co3-release-command-a-plus-addco4-release-command-a-plus-addco5-release-command-a-plus-add
    • cohere-command-a-familyRejected by two rubrics in round 5, and the only claim in this run to draw a substantive objection from editorial. The family record borrowed its description from the Command A release row and carried the modality of a later sibling: it listed multimodal-generalist because Command A+ accepts images, while the same table lists Command A itself as text-only, and it adopted the marketing word "excelling" from release copy into family-level prose. Editorial held that family and release were not kept distinct. That is a real modelling defect rather than a sourcing gap, and it is recorded here rather than patched, because correcting it would need a fourth panel on a claim whose release dependency had already failed.Blocked by co5-family-command-a-add

    Accepted by the panel, then dropped

    • mistral-ai-organization-and-familyThe Mistral AI organization, the Mistral 3 family, the publisher and the announcement source were all accepted unanimously, 3-of-3. They were still not applied: model-tree.ts places a creator in a branch only when it has at least one release, and drops families holding zero releases, so an organization and family without an accepted release would have added records that render nowhere and read as data while showing nothing. This is the same defect as issue #441.Blocked by mistral-large-3
    • cohere-organization-and-sourcesThe Cohere publisher and three sources were accepted unanimously across rounds 3 and 4 and were carried into round 5 with their original verdicts rather than re-reviewed. The organization claim was rewritten in round 5 to rest on quotes that state the creator relationship directly and was accepted by consistency and editorial, but provenance rejected it because no quote classifies Cohere as a "company" and the sentence naming the Command models resolves its subject only from surrounding context. Nothing was applied: with no accepted release the creator renders in neither branch, and applying the sources alone would leave them orphaned, which the validator forbids.Blocked by cohere-command-a-plus-05-2026

    Verification date deliberately held back

    • microsoft-mai-thinking-1No approved-origin page states an accessType or an input/output modality set for MAI-Thinking-1. Both are required fields on releaseSchema. The only availability statement found — "available in public preview on Microsoft Foundry" — describes a serving platform, not the model's access type, and treating it as one would collapse creator and platform into a single entity. Recorded as unknown rather than inferred.Blocked by tools/updater/profiles/origins/microsoft.jsonweb/src/data/schema.ts

    Sources conflict, so no value changed

    • xai-grokx.ai and docs.x.ai both now self-describe as "SpaceXAI LLC" — docs.x.ai is titled "SpaceXAI Docs" — so the naming conflict spans every origin approved for this creator rather than being contradicted by one of them. The profile records it as an unresolved ambiguity and it stays explicit rather than being smoothed over. Separately, no approved page names a Grok family, and x.ai/news returns Cloudflare 403.Blocked by tools/updater/profiles/origins/xai.json

    What this run does not prove

    • Nothing was applied to the dataset. The Others branch is unchanged and still renders no creators; this entry is the only edit in the pull request.
    • The counts here describe the final adjudicated bundles — the round-4 Mistral bundle and the round-5 Cohere bundle. Five review rounds ran in total, and the superseded bundles proposed more claims than the 12 counted above; they are published in the pull request body with every verdict and dissent verbatim, not summarised here.
    • A unanimous 3-of-3 bar means a single rubric decides the outcome, and provenance was that rubric in every rejection but one. Three provenance reviewers on different model families rejected these releases across three rounds on shifting grounds, so this is not one model’s idiosyncrasy — but neither is it evidence that the underlying published facts are wrong.
    • Round 5 is the useful control. It answered round 4’s stated objections with better evidence from the same approved origin — an explicit Status column, a specification table giving licence, size, context length and modalities — and the claims still did not carry, because the surviving objections are about the dataset’s own vocabulary rather than about the sources. If mapping a creator’s "Live" to the schema’s "current" is an unsupported inference, then no release record can ever clear a 3-of-3 provenance bar, whatever the evidence.
    • A withheld creator is a real outcome of the policy, not a failure of the sources. Every fact these releases rest on is published by the creator on an approved origin.

    Follow-ups — proposed, not fixed

    • The three rubrics have no shared rule for whether dataset-vocabulary fields — status, categories, accessType, license.osiApproved — are assertions about the world needing their own verbatim quote, or encodings of facts already quoted. Round 5 isolated this as the single blocking question: the evidence objections were answered and the claims still failed. Settle it before the next long-tail attempt, because at present no release record can reliably clear a 3-of-3 bar.
    • Related and narrower: decide whether normalising a quoted "128K" into the integer 128000, and reading a release date from an announcement’s datePublished, count as inference. The committed dataset already does both — openai-gpt-4-1 takes its firstReleaseDate from its announcement — so the standard applied to new claims is currently stricter than the one the existing data was built to.
    • licenseSchema requires osiApproved whenever a license object is present, which makes an otherwise well-sourced Apache 2.0 licence unrecordable unless an origin stating OSI approval is approved. Either make osiApproved optional or approve such an origin; today the only honest option is to drop the licence entirely, losing a fact the creator publishes plainly.
    • Microsoft cannot be recorded as a model creator from its currently approved origins: microsoft.ai and the bounded www.microsoft.com research-blog path state no access type or modality set for MAI-Thinking-1. Either an additional creator-side origin is approved, or Microsoft stays absent; azure.microsoft.com and learn.microsoft.com are the wrong entity and must not be used to close the gap.
    • x.ai self-describes as "SpaceXAI LLC" across both approved origins while the profile and the wider world still say xAI. The conflict is recorded and unresolved, and until it is resolved the creator has no defensible canonical name.
    • x.ai/news returns a Cloudflare 403 to this fetcher, so xAI release announcements were never readable from an approved origin.
    • No Mistral context window is recorded anywhere in the dataset: the only page stating one redirects outside the approved allowed_paths for docs.mistral.ai.
  2. 2026-08-26-d5f729

    Data refresh 2026-08-26

    Scope requested: Every creator in organizations.json, plus the long-tail profile sweepPublished63 edits posted · 9 items withheld

    An agent-run, source-backed refresh under ADR 0003. Five scouts fetched and hashed 55 primary pages and proposed 71 claims; twelve blind reviewers cast 213 verdicts across three rubrics; four deterministic gates passed; 63 edits landed across five dataset documents and GitHub Pages deployed successfully. No human reviewed any part of it.

    1. PreflightRanClean tree, gh authenticated, no unmerged pull request from a previous refresh.
    2. ScoutRanFive bundles, one per creator profile plus a long-tail sweep. 55 page snapshots hashed.
    3. ReviewRanTwelve reviewers, three rubrics over four bundles. The long-tail bundle had no claims to review.
    4. GatesRangate-evidence and gate-source-approval before applying anything; gate-dataset, npm run validate, and gate-scope after.
    5. PublishRanPR #302 opened with the full evidence trail and merged by GitHub on a green web-ci.
    6. DeployRanPages deployed main @ 0e9867b successfully; the live site serves the new records. No revert needed.

    What was found

    Scouts
    5
    Pages fetched and hashed
    55
    Claims proposed
    71
    Claims per creator bundle, with the review threshold its profile set
    CreatorPolicyThresholdClaims
    openaipilot2-of-314
    anthropicpilot2-of-318
    google-deepmindpilot2-of-319
    metapilot2-of-320
    long-tail sweeplong-tail3-of-30
    What those claims proposed to do
    KindCountEffect
    Add25New records.
    Change15Field corrections and counter updates.
    Unchanged29Re-verified against a primary source; no value changed.
    Conflict2Primary sources disagree; both sides recorded, no value changed.

    Not covered

    • Any creator outside the four pilot profiles and the long-tail catalogue.
    • www.llama.com content, deferred because the host now redirects to an unapproved origin.
    • The long-tail sweep returned zero claims. That is an honest zero, not a skipped stage — it was gated like every other bundle.

    What was evaluated

    Reviewers
    12
    Verdicts cast
    213
    Accepted by panel
    71
    Rejected by panel
    0
    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.
    CheckScopeExitResult
    gate-evidence5 bundles0PassThree rubrics present per claim, no duplicate votes, no empty rationale, every claim at its profile threshold.
    gate-source-approval5 bundles0Pass134 citations against anchor 13276b0, selected by merge-base with refs/remotes/origin/main. 20 inherited sources, 17 proposed, 16 approved origins, no new host admitted.
    gate-datasetworking tree0PassThe dataset is internally coherent after the edits were applied.
    gate-scopebranch vs anchor0Passchanged: 5, outOfClass: [], empty: false. Exit 0 alone is ambiguous; the publish precondition is exit 0 and changed > 0.
    npm run validatebaseline and after0Pass400/400 tests, 0 errors both before any change and on the branch. The baseline was captured first so a pre-existing red main could not be mistaken for damage from this run.
    web-ciPR #302—Pass400/400 tests, 0 errors. This is the check branch protection requires before a merge.
    skills-cinot requiredPR #302—FailRed, and not a required check, so GitHub merged on web-ci alone before it could be acted on. Filed as #304: gates.test.mjs pins TODAY to 2026-08-25 but asserts it against the live dataset, so any same-day refresh fails it.

    Applied over a recorded dissent

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

    • openai-add-gpt-5-6-cyber-releaseEditorial and entity boundariesreleaseDate was taken from an RSS pubDate, which dates the announcement item rather than the release.
    • anthropic-unchanged-source-opus-5-announcement-checkedProvenanceThe quote supports the API identifier but not the pricing half of the same claim.
    • meta-muse-video-releaseEditorial and entity boundariesA model described as "coming soon" was given a concrete day-precision releaseDate.

    Posted 63 edits

    63 edits across 5 documents, a net change of 24 records.

    Dataset documents this run changed
    DocumentBeforeAfterWhat changed
    sources.json476114 sources added.
    families.json1112One family added: meta-muse.
    releases.json2231Nine releases added.
    organizations.json44One field corrected; no record added or removed.
    usage-observations.json22Both readings updated; no record added or removed.

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

    Records added

    Not posted 9 items

    Accepted by the panel, then dropped

    • openai-add-source-deprecationsAccepted by the panel, then dropped by the chair. validate.test.ts enforces that an unreferenced source is dead provenance, and no reviewed claim wired openai-deprecations into any record's sourceIds. Adding that edit would have meant authoring dataset content no reviewer saw.Blocked by validate.test.ts
    • openai-change-gpt-5-status-deprecatedDropped with its source. Keeping the status change alone records "deprecated" against sourceIds that do not state it — every supporting quote comes from openai-deprecations.Blocked by openai-add-source-deprecations
    • openai-change-gpt-4-1-nano-status-deprecatedDropped for the same reason as the GPT-5 status change: its only supporting quote comes from the source that could not be added.Blocked by openai-add-source-deprecations

    Verification date deliberately held back

    • releases/anthropic-claude-haiku-4-5The field was re-read and unchanged, but validate.ts requires a model-fit statement to be at least as fresh as the facts it reads. Bumping verifiedAt would strand a dependent statement, so it stays at 2026-08-15. Under-claiming a verification is honest; over-claiming is not.Blocked by fit-claude-haiku-4-5-release-currentfit-claude-haiku-4-5-family-legacy
    • releases/anthropic-claude-mythos-5Held at 2026-08-15 for the same freshness rule: nobody re-derived the dependent fit statement over the re-read field.Blocked by fit-claude-mythos-5-limited-availability
    • families/meta-llama-4Held at 2026-08-15 by a test-suite coupling rather than a data fact: model-fit.test.ts injects fixtures pinned at 2026-08-15 over the real seed records. Tests are not dataset documents, so this run may not edit them.Blocked by model-fit.test.ts

    Sources conflict, so no value changed

    • families/anthropic-claude-4-5.statusTwo Anthropic pages disagree and neither changed a value. The models overview lists Claude Sonnet 4.5 under "Legacy models (still available)"; the deprecations page lists claude-sonnet-4-5-20250929 as Active, retirement not sooner than 29 September 2026. Conflicting data stays explicit rather than being smoothed over.
    • releases/google-gemini-3-5-flash-lite.statusThe Gemini 3 developer guide states "All Gemini 3 models are currently in preview"; the platform docs give gemini-3.5-flash-lite launch stage GA, released 21 July 2026. Recorded with both sides quoted; no value changed.

    Source refused by the approval gate

    • www.llama.comThe host now redirects to developer.meta.com, an origin neither the committed dataset nor any profile catalogue stands behind. The Meta scout deferred it rather than citing it, so the Muse claims rest on ai.meta.com instead. Admitting a new host is a human decision.Blocked by gate-source-approval

    What this run does not prove

    • The gates catch malformed, impossible, unreferenced, and boundary-violating data. They do not catch a well-formed claim that is simply wrong.
    • Per ADR 0005 the evidence gate checks the form of a citation, never its remote content. A gate pass attests that the paperwork is complete, not that the page still says what it said.
    • The three reviewers share a failure mode — three instances of one model family reading the same page — so a source that is itself wrong can carry all three.
    • Auto-merge was armed correctly, but because skills-ci is not a required check GitHub merged on web-ci alone while skills-ci was red. No gate was skipped and no threshold lowered; it does mean an unattended run cannot currently stop on a non-required check going red.

    Follow-ups — proposed, not fixed

    • A scout proposing a status change must also propose the sourceIds change that carries its source. This cost three panel-accepted OpenAI claims.
    • Fit-statement freshness blocks legitimate verification bumps on anthropic-claude-haiku-4-5, anthropic-claude-mythos-5, and meta-llama-4.
    • www.llama.com now redirects to developer.meta.com, a new host that needs a human decision on the approved-origin list.
    • Consider making skills-ci a required check, so a red non-required job can stop an unattended merge. Requiring a check is a branch-protection setting, so it is an owner action rather than a change this program can make.
  3. 2026-08-25-902306

    Data refresh 2026-08-25

    Scope requested: Every creator in organizations.json, plus the long-tail profile sweepStopped, published nothing0 edits posted · 1 item withheld

    Stopped at preflight and published nothing. ADR 0003 precondition 2 was still open — the approved-source binding enforced on the proposal-only path had no equivalent in the publishing skill set — and ADR 0003's guardrail is that the automation does not run until the skill set is corrected. No dataset change, no branch, no commit, no pull request, no deploy.

    1. PreflightRanModelTree checkout, clean tree, gh authenticated, no open pull request from a previous refresh. Preconditions were then checked live and one was open.
    2. ScoutNot runZero pages fetched, zero claims extracted. The publisher was not authorised to run.
    3. ReviewNot runNo bundle existed to review.
    4. GatesNot runNo bundle existed to gate. A read-only dataset health check was run separately and is recorded below.
    5. PublishNot runNothing was applied, so there was nothing to publish.
    6. DeployNot applicableNothing merged, so nothing deployed.

    What was found

    Scouts
    0
    Pages fetched and hashed
    0
    Claims proposed
    0

    Not covered

    • Every creator in scope. Scouting never started, so no creator was checked for new releases.
    • Because scouting did not run, this run is not evidence that the dataset is current. It says only that the publisher was not authorised to run.

    What was evaluated

    Reviewers
    0
    Verdicts cast
    0
    Accepted by panel
    0
    Rejected by panel
    0
    Deterministic gates and required checks — 1 of 1 check passed. Exit 0 is a pass; exit 2 means the gate could not run and is never treated as one.
    CheckScopeExitResult
    gate-datasetnot requiredcommitted dataset, read-only health check0PassRun standalone rather than as a run gate: passed: true, failures: []. Counts at the time — sources 47, publishers 8, organizations 4, families 11, releases 22, usageObservations 2, usageSyntheses 0, modelFitStatements 7, modelFitEvidenceGaps 3.

    Posted 0 edits

    Nothing reached the dataset. No branch, no commit, no pull request.

    Not posted 1 item

    Blocked by policy before it could run

    • entire-runADR 0003 precondition 2 was open. gates.py enforced an approved-source binding per claim; the publishing skill set's gate-evidence.mjs never inspected sourceId against any approved catalogue, and gate-dataset.mjs resolves sourceIds only referentially against a file the refresh may itself patch. On that rule the publishing path was strictly more permissive, which stops the automation.Blocked by ADR 0003 precondition 2#167

    What this run does not prove

    • This run did not find "nothing to do". It found work for a human, which is why its summary issue was left open until the blocker was closed.
    • The read-only health check shows the published dataset was internally coherent. It does not show that it was current or true — gate-dataset has no network and does not check that a source still says what it said.

    Follow-ups — proposed, not fixed

    • #167 — the blocker: the publishing gates must refuse a run that adds a source record to sources.json and cites it in the same change.
    • #205 — gate soundness is emergent across the set rather than provable gate by gate.
    • #210 — gate-scope reported green on a committed out-of-class change when run with no --base.
    • #209 — no gate input self-reported by the subject of the gate may have a default.
    • #168 — nothing detects drift between the Python and JavaScript gate implementations.