Every DOI added under content/ resolves and the resolved titles/venues match what the source records claim (Newson 2003, Denevan 1996, Stenborg 2018, Lombardo 2025, Piperno 2024, Levis 2017 + the McMichael technical comment, Schaan 2000, Arroyo-Kalin 2021, Lopes 2024 — whose record even discloses a prior wrong title corrected against the DOI check). Every URL marked FAILS is a 403 publisher bot-wall or an unreachable journal host, and in each case the source record itself already says the publisher page is bot-walled and names the workaround copy actually read (Wayback, UvA/Deep Blue/Bournemouth repositories, PMC), so the failures are consistent with the diff's own provenance rather than evidence of fabrication; the Pärssinen 2026 Nature article postdates my knowledge but resolves, and temporal humility applies. Verification labels look honest: the Carvajal primary text and Hancock's book are marked `unverified` with explicit instructions not to lean on their wording, and the Xingu dating rests on a news feature that is labeled secondary and flagged in both the claim and the assessment. The assessment overlay is stamped with runId/model/date/promptVersion, `humanReviewed: false`, `role: draft`, so no unratified draft is dressed as a ratified or human judgment (§3.15, §4). Negative and undermining evidence is first-class (AMZ-C013, C014, C009, the McMichael comment), evidence direction is set on every record, source statements are kept separate from `editorInference` (§3.5, §3.9), and the chronology qualifier AMZ-E003 explicitly refuses to spend AD 100–300 grandeur as contact-era corroboration (§3.2, §3.4). The founder's funding of the central 2026 study is disclosed on the source record and in the article body, and the case carries that paper's population model at its own least favourable sensitivity (AMZ-C005), which is the honest handling rather than a concealment; the Hancock genealogy claim carries diagnosticity `low` with an explicit statement that advocacy history is not evidence in either direction (§7). AGENTS.md is untouched; the IMAGE_STYLE.md and generator edits are non-normative additions. Some claims (AMZ-C011, C014) bundle several instances of one proposition, which is ordinary imperfection rather than a §3.2 breach.

The PR adds a new case with claims, evidence, sources, research agenda, images, and an AI draft assessment, all in new files; there is no edit to AGENTS.md or other governance code. Claims are clearly atomic, with separate Evidence records that distinguish sourceStatement from editorInference, and each Evidence has an explicit direction (supports/qualifies) and strength, matching the epistemic model. Provenance is careful: each source has verification labels (ai_verified/unverified) and detailed notes about how it was accessed (Wayback, repositories, bot walls), and no locator appears invented or self-contradictory; the citation tooling confirms all DOIs/URLs resolve where expected, including CC-licensed images. The overview and assessments explicitly frame uncertainty and model dependencies rather than presenting single truth scores, and the AI assessment file is labeled as a draft, unreviewed, with runId/model/date, respecting the append-only AI-judgment rule. Sensitive/confidential material is not used; all content comes from published papers, open web pages, or CC/Wikimedia media with licenses and credits stated. No existing assessments or records are silently altered, no provenance is removed or weakened, and no checks or constitutional safeguards are adjusted; this is an additive content change that adheres to the project’s epistemic and transparency requirements.

The change perfectly adheres to the constitution's epistemic and formatting rules. The AI-generated assessment is explicitly labeled as an unreviewed draft with proper run and prompt metadata (§3.14, §3.15). Citations are fully real: although some publisher URLs returned expected HTTP 403s against the scraper, their corresponding DOIs successfully resolved to Crossref metadata matching the exact claimed titles and authors (§3.8). The distinction between source statements and editor inference is rigorously maintained across the evidence ledger (§3.6). The integration of founder-supplied context correctly treats it as workflow provenance, while anchoring all substantive claims to verified public sources.

The diff adds a new case with claims, evidence, sources, a draft assessment, and art; it does not touch AGENTS.md or weaken §3 checks. Source records use honest labels (unverified for the unfetched Carvajal primary and Hancock book; ai_verified elsewhere, with Wayback/repository/OCR workarounds noted), evidence records split sourceStatement from editorInference and mark direction, and the overlay assessment is stamped role: draft / humanReviewed: false. Mechanical checks show added DOIs resolving with titles that match the records (Lopes 2024’s title now matches the DOI after an admitted correction); publisher 403s are access walls on identifiers that still resolve, not missing citations. The 2026 Nature paper and Commons Fig. 2 URLs resolve (HTTP 200); unfamiliarity with that paper is not fabrication evidence. Compound featured claims and a secondary Xingu news citation are ordinary imperfections already caveated in-record, not paradigm violations.

Every DOI the diff adds resolves to metadata matching its source record — Denevan 1996 (Bluff Model), de Souza 2018, Lombardo 2025, Piperno 2024, Levis 2017 plus the McMichael Technical Comment, Schaan 2000, Stenborg 2018, Rodrigues 2015, Lombardo 2011, and both Revista de Indias DOIs (the .555 Crossref title 'Patters of Indian Depopulation…' matches Newson modulo a registry typo) — satisfying real-citations-only (§3.7, §4); the FAILS are uniformly HTTP 403 bot-walls (Wiley, Science, T&F, SciELO, Pulitzer, grahamhancock) whose corresponding DOIs resolve, and the diff's own verification notes already document those walls and the Wayback/repository workarounds used, which is honest labeling rather than fabrication. Verification labels are honestly graduated: the Carvajal primary text and Hancock's book are marked unverified/background-only and never carry load-bearing wording, while ai_verified is claimed only where content was checked, with one caught-and-corrected invented title (SRC-LOPES-2024) disclosed in the record itself — evidence the checks worked, not that they failed. The assessment overlay is stamped runId/model/date/promptVersion, humanReviewed: false, role: draft, and nothing presents it as ratified (§3.15, §4); claims are stamped ai_extracted with per-record provenance, editorInference is separated from sourceStatement (§3.9), mixed effects are split (E012/E013), negative evidence and chronology qualifiers are first-class (§3.11), and uncertainty is calibrated in plain language (§3.13). The founder-funding disclosure quotes the published paper's own public funding statement and displays the conflict prominently rather than hiding it — no confidential material is cited, since the founder-supplied brief was used only as discovery input with every load-bearing cell re-verified against public sources. Hancock's genealogy claim is framed neutrally at diagnosticity low with symmetric treatment of mainstream skepticism, consistent with the geo-project framing constraint. AGENTS.md is untouched, and the script change is additive art configuration that does not weaken any check; residual nits (Arroyo-Kalin 2020/2021 online-vs-issue year, PR body saying '20 sources' where 21 records exist) are ordinary imperfection, not violation.