The diff touches only pipeline code, protocols, docs and tests — no content/ records, no citations (the verification section finds none), and AGENTS.md is untouched. The substantive change (verify v6: a per-claim `bearing`, splitting a record whose claims bear in different directions into one record per direction, each stamped with model/runId/promptVersion/date/reason in `readerActs`, the drafter's `origin` preserved) tightens §3.5 direction honesty and §3.14 provenance rather than loosening it, and the refusal path (all-null bearing ⇒ record refused with the reader's reason) is preserved. Upgrading a fetched source from `unverified` to `ai_verified` with a note naming the date and run, keeping the drafter's original note verbatim, is an honest label under §4/§3.8 — it stops evidence leaning on text labelled as unread. Two things I would want answered on the record, neither rising to a violation: (a) `bareQuote` now strips `.,;:!?…` at both edges, so the test asserts a span quoted with `!` matches source text ending in `.` — the words inside stay exact, but terminal punctuation is no longer faithful, and I would prefer the strip be symmetric-tolerant rather than punctuation-agnostic, or at least documented as a known relaxation of the verbatim check (§3.8, §3.15); (b) the new source-label note asserts "its quoted passages checked" for any successfully fetched `unverified` source, including a `background` source that nothing quotes — the wording should be conditioned on passages actually having been checked, or it misstates what was verified. Neither the case limit (`--max-cases`) nor the `correctionBlocker` null/absent-field equivalence dodges a §3.15 check; both are recorded in DECISIONS.md with dates and reasons.

The revised `bareQuote` deliberately strips all leading and trailing punctuation, and the new test positively requires `quoteOccurs("may not have occupied all of Amazonia!", source-without-that-exclamation)` to return true. An invented exclamation mark can alter tone or meaning, so this goes beyond accommodating citation-bracket placement and weakens the exact-quotation check required by §§3.7, 3.8, and 3.15. The bearing safeguard is also not fail-closed: although v6 requires one entry per named claim, neither `VERIFY_SCHEMA` nor `applyReader` rejects missing, duplicate, or foreign claim IDs, and omitted claims silently retain the drafter's direction. Fix the quote matcher so it cannot accept substituted punctuation, and validate that a non-null bearing contains exactly one entry for every named claim before applying it.

The changes implement verify protocol v6, refining evidence direction and claim splitting without compromising epistemic rigor. Handling bearings per claim directly serves §3.5 by splitting multi-claim records that bear in different directions rather than applying contradictory directions across linked claims. Upgrading source verification status to ai_verified when the verifier successfully fetches and checks the text maintains honest verification labels under §4. Normalizing edge punctuation on quoted spans preserves verbatim word matching while preventing false negative mismatches from citation brackets. Detailed provenance stamps (readerActs) record who changed what, when, and why, strictly adhering to §3.8 and §3.14.

The diff is pipeline/protocol only (no content/ ledger records, no AGENTS.md edit, no run account). Verify v6 requires a per-claim bearing and splits multi-direction records into one record per direction with the same quote, keeping the drafter origin and stamping each change in readerActs (model, runId, promptVersion, date, reason), which tightens §3.5 rather than weakening it. Edge-punctuation stripping in bareQuote is limited to span edges with words still exact (tests reject altered wording), fixing false quote failures without inventing text. Marking a source the verifier actually retrieved as ai_verified, while retaining the drafter’s note, is honest verification labeling under §3.15/phase-1. The --max-cases 2 sitting cap and inbox commit path are operational so the panel can see the PR; they do not loosen epistemic gates. Docs and the archived v5 protocol record the panel findings and the change.

No AGENTS.md edit, no content added, no citations introduced, and no run records to audit — the change is pipeline/protocol work, so fabrication and confidence risks are minimal. The three substantive fixes each preserve or strengthen epistemic checks rather than dodge them: the per-claim bearing split stamps every direction change, dropped link, and split as the reader's act with model/runId/promptVersion/date/reason (§3.9, §3.14 provenance preserved); upgrading an unreachable source to `ai_verified` only when the verifier itself retrieved and checked its text makes the §4 label more honest, not less, and keeps the drafter's note; and `--max-cases 2` exists precisely so panel seats can actually read the PR, reinforcing §3.15 ratification rather than weakening it. The `bareQuote` edge-punctuation change is the closest call under 'no weakening of the checks': it loosens the verbatim matcher at span edges, and because both '.' and '!' (and '?' and '…') are stripped, a quoted question could in principle pass as a statement — but every word inside the span must still occur exactly in the source, the adversarial reader's `quoteInContext` check independently tests negation and attribution, and the change is tested against the real text, so this is a fix for false rejections (a quote closed with its period against a citation bracket) rather than a dodge of a gate. The `correctionBlocker` change aligns the blocker with the writer's own semantics for absent fields and does not open a path past a genuine mismatch. I would flag as a review note that leading/trailing '?!…' substitution in `bareQuote` is broader than the stated motivation, but it does not break a specific constitutional rule.