No AGENTS.md edit; no content/ records change and no citations are added (verification tool found none). The diff adds a bounded PDF reading path that tightens rather than weakens the checks: Poppler-based container inspection before any paid call, hard 10 MB / 60-page limits with no silent truncation, a new requestAddressed relevance boolean, and — for PDFs — a mandatory independent checker confirmation of the quote against a separately rendered page image, enforced in code (`passes` requires quoteSupported/locatorSupported/pageImageHash) and in the Zod schema (`ResearchProposalSchema` rejects v6+ PDF passages lacking a rendered-page receipt). Provenance handling is honest: textHash is set to null for scans rather than fabricating a mechanical text-layer match, the locator explicitly says 'file page count, not printed pagination; AI page reading', and the source verificationNote states this is not human verification (§3.8, §3.9, §4). The failed pilot is preserved rather than buried: the v5 proposal with the wrong PDF page 31 remains in intake but carries two appended `research-adoption: invalid` decisions naming the scope failure and the page error, so nothing unratified is presented as a finding (§3.14, §3.15). Discovery/no_change receipts explicitly state that a reader's absence claim is not verified negative evidence (§3.11), and all new intake records are stamped with runId, model, promptVersion and date.

The change strengthens provenance checks by bounding PDFs, hashing the retrieved file and rendered page, requiring an explicit physical-page check, and labeling PDF verification as AI reading rather than mechanical text matching or human verification. The known v5 wrong-page result is preserved as a proposed draft, then explicitly invalidated by two append-only research-adoption records, including the corrected page description and page-image hash; it is not installed as canonical evidence or presented as ratified. The v6 no-change result remains a model explanation rather than negative evidence, and the documentation expressly says the comparison remains unresolved. The reader also checks request relevance, separates source statement from inference, records dependency notes and independence groups, and leaves the ordinary cross-vendor publication gate intact.

The pull request adds bounded PDF source reading capabilities alongside mechanical Poppler page rendering and strict relevance and physical-page checks. No files under `content/` are modified, and no canonical claims, evidence records, or assessments are published. The durable intake records accurately preserve the full audit trail of pilot runs, including schema validation failures and explicit adoption invalidations due to out-of-scope observations and page mismatches, fully complying with provenance and append-only versioning rules.

The diff adds bounded PDF intake (10 MB / 60 pages, Poppler inspection, inline file + independent rendered-page image, null textHash, explicit AI page-check receipts) and a request-relevance gate without touching AGENTS.md or any canonical claims, assessments, or articles. Failed pilot runs are preserved as failed/invalid/no_change intake (schema miss, out-of-scope observation, wrong page 31 confirmed by later page-image review); the v5 proposal is not adopted and is labeled AI-extracted/ai_verified with honest non-human caveats. Source checks stay separate from the cross-vendor publication arbiter, strengthening rather than weakening §3.15 gates. No fabricated locators are installed as ratified evidence; documentation and DECISIONS.md accurately record the pilot costs, rejections, and founder steering.

This PR adds infrastructure (bounded PDF input, page-image verification, relevance checking) and intake receipts, not published content — no canonical claims, assessments, or articles change, so the publication-ratification rules (§3.15) are not implicated; the one v5 proposal that advanced was explicitly rejected as invalid and preserved as a failed pilot, never presented as a ratified or human finding. The v5 failure (both models confirming a wrong page locator) is transparently recorded in append-only intake decisions with provenance (model, runId, promptVersion), and the no_change outcome is explicitly labeled an unverified reader explanation rather than negative evidence, consistent with §3.11/§3.13. The verification labels are honest: PDF passages get textHash null, 'ai_verified'/'AI page reading' labels, explicit 'not human verification' limitations, and the new schema (§3.8) forbids inventing text-layer hashes while requiring quoteSupported/locatorSupported confirmation plus a rendered-page receipt for v6+ readings. The apparent pageImageHash exemption for the v5 receipt preserves an append-only historical record rather than weakening checks — the v6 superRefine rule still enforces the rendered-page receipt for all new PDF readings, and the diff strengthens rather than weakens the mechanical checks (strict schemas, fail-closed PDF bounds, hash binding without byte exposure). No citations are added, so there is nothing to fabricate; the Met URL and search-lead URLs are labeled discovery aids, not verified evidence.