The diff touches only the arbiter gate script, its shared core, a test, and the decisions log — AGENTS.md is untouched, and no content/ citations are added (verification found none). The effect on the §3.15 checks is a tightening, not a loosening: the wholesale epoch-5 reset is withdrawn (GATE_EPOCH returns to 2026-09-09T15:00Z, an earlier window start, so more merges fall inside the count), the loop's own merges (#245, #257) are put back into the autonomous tally, and the exclusion is narrowed to nine specifically enumerated founder-directed squash commits, each stamped with hash, PR number, and the direction and date it answered. `splitMergeLanes` gains an explicit, defaulted `declared` parameter that is exactly equivalent in force to the pre-existing `Supervised-by:` trailer, and the added test pins both the declared and the default-counts behaviour, including that unmarked merges still count when no declaration is passed. Recording founder directions given in session, with date and channel, is expressly permitted by the §3.15 operating-record clause. I cannot independently confirm the nine commit hashes or PR numbers, but unfamiliarity is not evidence of fabrication and nothing in the material is internally inconsistent — the list is closed, auditable, and stated not to grow because new construction commits carry the trailer. The change also answers an outstanding review note (#261) on the record, as the lone-objection rule requires.

The change replaces a wholesale epoch reset with an auditable, per-merge classification: nine exceptions are identified by full commit hash, PR number, dated founder direction, and session or review-note context. Restoring epoch 4 causes the loop’s own merges (#245 and #257) to remain counted, so the rate-limit check is not evaded in the manner identified by review note #261. The core function still defaults unmarked and undeclared commits to the autonomous lane, and the added test covers declared, trailed, and unmarked cases. The withdrawal and rationale are appended to docs/DECISIONS.md, and neither AGENTS.md nor the constitutional checks are weakened.

The change withdraws the wholesale epoch reset (epoch 5) and replaces it with an auditable, per-commit hash declaration of nine founder-directed landings that lacked the `Supervised-by:` trailer. This restores the loop's autonomous merges (#245, #257) to the rate-limit accounting rather than wiping them out, accurately reflecting true lane usage. The update adheres to §3.14 and §3.15 by documenting the decision in DECISIONS.md and adding regression testing without weakening arbitration checks.

The diff restores GATE_EPOCH to epoch 4, adds an explicit SUPERVISED_DECLARED hash set for nine named squash commits, and threads that set into splitMergeLanes so trailer-less founder-directed landings are classified per-merge rather than by wiping the window. That answers review note #261 without editing AGENTS.md, without touching content citations, and without loosening the autonomous rate-limit itself: unmarked merges still count, trailed merges still exclude, and the declared list is documented as closed because future operator commits carry the trailer. docs/DECISIONS.md records the withdrawal, the review-note trigger, and the founder-direction rationale with date; the unit test covers declared, trailed, and unmarked cases. No fabrication, confidence exposure, or check-dodge is visible in the complete diff.

This is an operational repair to the merge-lane gate, not a content change, and it strengthens rather than weakens the checks: where the withdrawn epoch-5 bump (#260) cleared the weekly count wholesale — erasing even the loop's own merge #245 — this change restores epoch 4, counts the loop's merges (#245, #257) toward the ten, and replaces the wholesale reset with a per-merge, auditable declaration list naming each commit by full hash with its PR and the founder direction it answered, exactly what review note #261 requested. The nine PRs and their founder-directed character are corroborated by the prior, already-present comment block in scripts/arbiter.mjs (visible in the diff's removed lines), which lists the same nine PRs with matching descriptions, so I see no positive evidence the hashes or directions are invented, though I cannot independently verify git hashes from the packet — if a hash were wrong, that would be a provenance defect the operator should fix on the record. The declaration semantics match the existing `Supervised-by:` trailer mechanism (merged messages are immutable and cannot take a trailer retroactively), the founder directions are stamped with date and channel per the §3.15 operating-record rule, DECISIONS.md records the withdrawal and rationale, and AGENTS.md itself is untouched. The test covers declared, trailed, and unmarked cases, and the stated intent that the list not grow (operator commits now carry the trailer) is the right guard against future evasion.