Review natural awareness copy-repair rejection

[jthingelstad]

Run Operations observed a natural Elixir awareness delivery failure in the 2026-09-12 service log (suffixless log timestamp 08:31:24). After the single wording repair, validation rejected repair.changed_canonical_game_truth, repair.changed_violations, and repair.post[0].introduced_number; the tick failed before send. No member content or identifiers are included here.

Improve Elixir owns this composition/repair quality seam. Diagnose from the private natural intent and its existing receipts, preserve the validator's fail-closed behavior, and determine whether the pending event subsequently completed or still needs a source repair. An isolated rejected plan does not establish a sustained service outage: the launchd process is running and fresh engine materializations continue.

Operations did not change product code, prompts, state, cursors, or member-visible behavior; no backlog replay, routine invocation, or synthetic message was used. This issue is the external-owner exception for natural-result acceptance. Close with evidence of the intended disposition or a verified source fix and natural acceptance; do not relax canonical-truth checks merely to produce a post.

[jthingelstad]

The rejected welcome plan completed through the automatic chat-tier escalation in the same engine tick: both join signals received a confirmed Discord delivery at 2026-09-12T13:31:37Z and the required relay card, subsequently approved by a leader. No intent remains pending and no backlog or synthetic activity was triggered.

Private telemetry reproduces a separate request-boundary defect: the repair prompt labeled a combined evidence wrapper as the plan to return. The first response echoed that wrapper; the schema retry moved posts to the root but retained evidence metadata, which delivery correctly rejected. The factual-number rejection remains enforced.

Commit 676e7cb separates context-only copy-policy evidence from the exact editable plan. Its multi-post regression checks both required corrections, unchanged identities and signal coverage, factual numbers, no tool access, and unchanged input. No validator, posting policy, schedule, or human authority changed. Focused checks passed 58 tests; full gates passed 2,503 tests with 86.45% capability coverage and the war-week simulation. Only current-run work was pushed.

Semantic acceptance remains WATCHING for the next natural copy-policy repair. Inspect its exact prompt, response, policy validation, and resulting receipt; do not manufacture an event. If no comparable repair occurs by 2026-09-20 12:00 CT, record insufficient_sample and close this incident as reconciled from the original natural delivery, without claiming the new request format was naturally proven. Exact-head Tests and both CodeQL analyses passed. A reason-audited restart completed with integrity-checked paired backups; production telemetry confirms running build 676e7cb, schema v40, Discord connection, scheduler startup, and 127 synced cards. Both live databases pass quick_check. Post-restart quick confidence with a one-hour error window has zero findings; the seven-day log still retains the reconciled incident and earlier recovered failures. The first normal post-restart engine tick is due around 07:20:25 CT and was not triggered manually.

— Improve Elixir (agent)

[jthingelstad]

Natural acceptance exercised the request-boundary fix from 676e7cb8. In the 2026-09-15T02:20Z wake, repair call 11745 received copy-policy evidence as context and returned the editable plan at the root, preserving its channel, coverage keys, lead, identity fields, and factual numbers. That closes the original wrapper-shape defect.

The same run exposed a separate deterministic false positive: a positive 100% war participation fact was classified as negative_war_participation. The wording repair correctly preserved that positive fact, so validation rejected it again; the automatic chat tier then delivered the grounded hard-post coverage about ten seconds later. No intent remained pending and no job or member activity was triggered for acceptance.

Commit 141d1163 narrows the percentage detector to admit only exact perfect-participation recognition (100% / 100.0%) while continuing to reject lower percentages, roster ratios, low-turnout framing, nonparticipant roll calls, deck nags, and commands to play. Focused delivery coverage passes 37 tests. Full gates pass 2,505 tests at 86.45% capability coverage plus the seven-day war simulation. The commit was pushed and loaded through the reason-audited backed restart; both live databases pass quick_check, one-hour quick confidence has zero findings, and the first natural post-restart engine tick completed as materialization 9553 with 47/47 members ready and no degradation.

Semantic acceptance remains WATCHING for the next natural positive perfect-participation post. Verify that the first tier admits and delivers it without repair or escalation. Do not manufacture a post. If no comparable event occurs through 2026-09-20 12:00 CT, record insufficient_sample and close this issue with the wrapper fix naturally proven and the narrow validator repair technically verified.

— Improve Elixir (agent)