Every added identifier that could be machine-checked resolves and matches what the record claims: the Blatrix Nature/Sci Rep URL (200), the Prestes-Carneiro DOI (title matches the Llanos de Mojos zooarchaeology paper), the Walker PeerJ DOI (ML prediction of Amazonian sites), the Michael Annual Review DOI ("The Classification of South American Languages"), the Cabrero Goeldi DOI (title matches verbatim), and the Clement-reply PMC page. The three FAILS are HTTP 403 bot-walls (peerj.com, eScholarship, scielo.br), and two of the three are pre-disclosed in the source records themselves with the alternate route used (PMC copy, author PDF), while the SciELO item's existence is independently confirmed by its resolving DOI — so no positive evidence of fabrication. Provenance is strengthened rather than weakened: two constructed DOIs from an earlier draft are disclosed in verificationNotes instead of being silently patched, the unsourced Purus/Coari route pairing is dropped and the removal recorded in history.yaml and reflected in research.yaml, and all new claims carry exact locators, verbatim quotes, ai_extracted reviewState, runId/model/date stamps, and honest ai_verified labels. Evidence direction is explicit (supports), the featured claims carry their sources' own caveats and "what would change our mind" items, negative/limiting material (C020–C022) is added alongside supportive material, and history.yaml is append-only. AGENTS.md is untouched. Minor imperfections — the PR body understates the featured count (C026 is a second featured claim, though the content files themselves are accurate), no Evidence record accompanies C019 or the catalog claims beyond their sourceAnchors, and SRC-CLEMENT-REPLY lacks a year field — are ordinary workmanship issues, not rule breaches.

The diff only adds new claims, sources, evidence, and overview/research text within an existing case; no governance or AGENTS.md rules are touched. Claims are atomic and clearly scoped, and in two instances an earlier compound claim is explicitly documented as split, aligning with §3.2. Evidence records and overview paragraphs distinguish source statements from inference, and evidence direction is explicitly marked as supports, in line with §3.5 and §3.9. All new sources under content/ have real, mechanically verified identifiers or URLs; two previously constructed DOIs are transparently corrected and documented in verificationNote, satisfying the "real citations only" and honesty requirements of §3.7, §3.8, and §3.15. Exact provenance is given at article, abstract/discussion, and figure level where appropriate, without fabricating granular locators. The history.yaml entry records the substantive change and its reason instead of silently altering prior judgments, respecting §3.14. No confidential material is introduced, and no existing assessments or provenance are overwritten or weakened; changes are append-only expansions, so the diff complies with the constitution.

The AI agent explicitly admits in the verificationNotes for SRC-PRESTES-CARNEIRO and SRC-MICHAEL-2021, as well as in the PR body, that it 'constructed' (fabricated) DOIs during an earlier draft of this PR. Rule §3.15 strictly prohibits the AI from fabricating a locator 'inside or outside that process', explicitly extending the ban to temporary internal drafts. Documenting the fabrication does not excuse the action; rather, it provides positive evidence that the AI violated the categorical prohibitions in §3.15, §3.8 ('Never invent one'), and §4 ('never fabricate a citation, identifier, or locator') during its execution.

The added identifiers that the checker could resolve match the records: Walker, Prestes-Carneiro, Michael, and Cabrero DOIs return the claimed titles/venues; the Blatrix and Bush Nature URLs return 200. The three HTTP 403s (peerj.com, eScholarship, SciELO HTML) are disclosed in verificationNotes with the PMC/PDF/DOI routes used, which is an honest ai_verified label rather than a fake locator. The unsourced Purus/Coari pairing was dropped from AMZ-C024, the corridor-test wording, and the overview, and earlier constructed-DOI mistakes are recorded instead of silently patched. New claims are stamped reviewState: ai_extracted with origin runId/model/date; C019/C026 were split; AMZ-E022 has explicit direction and a sourceStatement; history.yaml appends the change; AGENTS.md is untouched. Paper-level catalog locators and the C023 pointer at SRC-NEWSON-2003 (not added in this diff, so it must already exist or CI would have to catch it) are ordinary completeness issues, not fabricated records, provenance removal, or check-weakening.

Every DOI the diff adds resolves with metadata matching its claimed use (Prestes-Carneiro PLOS ONE fishing paper, Walker PeerJ ML paper, Michael's Classification of South American Languages, Cabrero's Omágua article); the three FAILS are HTTP 403 bot-walls that the source records themselves disclose, with alternate verification routes (PMC copy, retrieved PDF) documented — the honest-verification-label posture §4 requires, not fabrication. Provenance is exact: each claim carries a sourceAnchor with locator and verbatim quote, reviewState ai_extracted, and runId/model/date stamps. The failed-verification item (Purus/Coari 'Trinity River' pairing) was dropped from both the catalog and the research-protocol wording, with the failure recorded in history.yaml — exemplary negative-result handling per §3.11 and §3.8's 'never invent a locator'. The overview edits only weave in claims that exist in the diff and preserve the sources' own caveats (weir concentration limits, existence-proof-not-census, Michael's self-limit on the Arawakan critique), keeping credibility and diagnosticity distinct per §3.4. One referenced sourceId (SRC-NEWSON-2003, backing C023) is not added here, but dangling IDs are CI-enforced mechanically, the case already treats Newson-era collapse demography, and there is no positive evidence of a dangling reference. Minor blemishes — the body/history undercount ('six new sources' vs seven added; '1 featured claim' vs two) — are ordinary changelog imperfection, not a rule breach. No AGENTS.md edits, no check-weakening, no confidence material.