The diff adds a new edition protocol (v8), archives v7, adds a DECISIONS.md entry, and extends the packet's claim index with an `evidence` count. AGENTS.md is untouched. No citations, locators, or quotes are added, and the citation tooling found no identifiers, so there is no fabrication surface. Substantively the change strengthens rather than weakens the §3 checks: it enforces §3.6 (a source anchor is not evidence), §3.9 (source statement vs inference, including in the steelman), and §3.13 (anchor-only claims graded `unresolved` at `low` confidence with the basis stated), while preserving §3.3/§3.11 selection duties (thin-evidence competing claims are still featured but graded honestly). The DECISIONS entry is stamped with date and channel ('founder direction in session, 2026-09-16'), consistent with the operating-record provision in §3.15. One small inconsistency: the PR body says anchor-only claims are graded 'at most provisionally_supported at low confidence' whereas the committed protocol and DECISIONS say `unresolved` at `low` confidence, never `supported` — the body appears stale relative to the stricter final text, which is the one that governs; I would ask that the PR description be corrected, but this is a description-hygiene issue, not a rule breach.

The packet now exposes a separate count of admitted Evidence records, preventing a source anchor from being treated as evidence contrary to §3.6. Protocol v8 requires anchor-only claims to be attributed to their source, graded unresolved at low confidence absent another disclosed basis, and textual inferences to be identified as proponents’ readings, which strengthens §§3.9 and 3.13. The change adds no citations, confidential material, ratification claims, or edit to AGENTS.md, and the dated decision entry records its stated supervision and channel rather than weakening the publication checks.

The change updates the edition protocol to version 8 and exposes the count of admitted evidence records per claim in the packet builder. This directly enforces §3.6 ('Sources are not evidence by themselves') and §3.9 ('Source statement versus inference') by preventing claims with only source anchors from being graded as supported or asserted as established facts in narrative drafts. The operational record in docs/DECISIONS.md complies with §3.15's governance for documenting founder directions, and AGENTS.md is untouched.

The diff only archives edition-v7, adds edition-v8, extends the packet claim index with an admitted-evidence count beside anchors, and logs a DECISIONS entry stamped with founder direction. It does not touch AGENTS.md, add citations or content records, or loosen any §3.15 check; v8 instead hardens §3.6/§3.9 by requiring anchor-only claims to be voiced as held on the anchor, graded unresolved/low unless a non-anchor basis is disclosed, and by writing textual inferences in the steelman as proponents' readings. No run account or omitted files; nothing in the visible change fabricates provenance or publishes confidence material.

The change tightens rather than weakens the epistemic checks: edition-v8 requires anchor-only claims to be told as held on their anchor (never in the page's own voice), graded `unresolved` at `low` confidence with a narrow disclosed-basis exception, and bars grading them `supported` — a direct application of §3.6 (a source is not evidence) — and it extends §3.9 to the steelman by requiring textual inferences to be written as the proponents' reading. The packet change merely exposes the admitted-evidence count beside the anchor count so the drafter can see the distinction, and the DECISIONS.md entry records the founder's session direction in the format §3.15's amendment permits, matching the adjacent prior entry's pattern. No AGENTS.md edit occurs (only protocol, packet, and decisions files), no citations are added, no runs are carried, and the diff is complete with nothing omitted. The only wrinkle — the PR body says 'at most `provisionally_supported`' while the protocol file actually mandates `unresolved` — is a body/diff discrepancy in the stricter direction, so the diff governs and the discrepancy costs nothing epistemically.