The diff touches only tooling, protocols, and docs: an arXiv PDF-first retrieval path in src/pipeline/fetch.ts, larger token ceilings for the splitter/judge, fuller rejection reasons in verify.ts, a new draft protocol v7 (with v6 archived), and DECISIONS/AUTOMATION entries. AGENTS.md is untouched, and the verification tool found no citation identifiers in added content lines (nothing under content/ changed). None of the changes loosen a §3.15 check: raising maxTokens from 4000 to 16000 fixes a mid-thought truncation failure rather than relaxing a gate; printing the whole failing span plus the source it was checked against strengthens the honesty of rejection reasons (§3.8, §3.14); draft-v7's rule that a research item may state no result or figure from a source the ledger does not hold tightens provenance discipline (§3.7, §3.8). The retrieval change keeps the abstract page as the record's key while reading the PDF text and discloses the substitution in `via`, which is honest provenance rather than a misattribution. The narrative claims in DECISIONS.md (run #266, seat findings, $3.42, the Amazon primaries) are internally consistent with the code changes and tests, and I see no positive evidence of fabricated locators or quotes; my unfamiliarity with the arXiv ids used only in test fixtures is not evidence of fabrication, and they are synthetic test inputs in any case.

The retrieval change preserves the arXiv abstract URL as the source key while explicitly recording the PDF in `via`, improving rather than obscuring exact provenance and ensuring drafter and verifier inspect the same text. The verifier now reports complete rejected spans and identifies the checked source text, without loosening verbatim-quote admission. Draft protocol v7 strengthens the ledger boundary by prohibiting research items from carrying findings or figures outside sourced evidence records, while retaining atomicity and source-versus-inference requirements. The diff neither edits AGENTS.md nor adds public content citations, assessments, or run records represented as ratified.

The proposed changes strengthen pipeline verification and provenance without altering AGENTS.md or weakening any checks. Upgrading arXiv retrieval to fetch the full PDF text aligns the verifier with what the drafter inspected, preventing false quote-matching rejections on abstract-only pages. Draft protocol v7 explicitly prevents research agenda items from introducing unsourced findings or figures outside of evidence records (§3.6). Un-truncating verification rejection strings improves inspectability and error diagnosis (§3.8, §3.15).

The diff only hardens pipeline and protocol behavior: arXiv abstract URLs now fetch the PDF under the abstract key with an explicit via stamp so drafter and verifier share text; rejection reasons print the full quoted span and name the source text; reader/splitter budgets rise; draft-v7 forbids research items from stating results or figures absent from the ledger. draft-v6 is archived, not silently rewritten. No AGENTS.md edit, no content/ citations, no run records presented as ratified, no confidence material. DECISIONS.md and AUTOMATION.md record the sitting faults and fixes without fabricating provenance.

The PR does not touch AGENTS.md (only docs, protocols, and src), so §3.15's amendment reservation is not implicated. Its four changes all tighten rather than loosen the epistemic checks: the arXiv retrieval fix makes drafter and verifier read the same text so verbatim-quote checks (§3.8, §3.9) are exercised against the actual paper rather than failing spuriously; the rejection-reason change prints the full quoted span and names the source text, which increases transparency of the verify gate rather than weakening it; the token-limit raise is a resource fix, not a check change; and draft-v7's rule that research items state no finding from a source the ledger does not hold strengthens the ledger against unsourced figures entering the public agenda. No citations are added under content/ (the verification section is empty), the arXiv identifiers in the new test are fixtures against a mocked fetchImpl and are not real-citation claims, and no confidence material, provenance misattribution, or unratified-draft-as-ratified presentation appears in the diff. The DECISIONS.md entry honestly records the sitting's failures as the loop's own, which is consistent with §3.14/§3.15's accountability requirements. Nothing here weakens a §3.15 check or dodges ratification — the sitting itself is being re-run under the fixed rules rather than hand-repaired, which is the constitutionally cleaner path.