Correct W38 story and report grounding gaps

[jthingelstad]

Decision: repair factual grounding for the next natural flagship cycle

Improve Elixir's 2026-09-14 read-only audit found two exact trust failures in otherwise successfully delivered W38 outputs. This is a durable audit and one bounded Jamie decision, not a dispatch ticket. No source, prompt, schedule, persona, leadership policy, stored member content, or natural job was changed or triggered.

Jamie decision

Approve a bounded factual-grounding repair for the next naturally scheduled weekly story and member reports?

The smallest scope is supplying canonical collection and comparable-stat evidence, adding regressions for the failures below, and enforcing the already-stated grounding bar. It does not authorize editing or deleting existing Discord messages, resending emails, resolving leader cards, changing posting worthiness/cadence, or implementing the deferred daily-war-relay suppression. If declined or deferred, preserve the current behavior and reassess against the incoming work Jamie mentioned on September 13.

Exact natural evidence

Unsupported first-of-rarity claim propagated into the clan story

• Player-stream events 1303 and 1304 record Mother Witch and Phoenix (both Legendary) on 2026-09-08; event 1319 records Sparky on September 9. Event 1362 records The Log on September 13. Its payload says only card ID, name, and rarity, not that this was the member's first Legendary. Current collection contains 14 Legendary cards.
• Brain loop 479 invented the first-Legendary characterization. Confirmed post 262 / Discord 1548696594394648790 and still-proposed relay R358 contain it.
• Natural memory synthesis at 2026-09-14T03:03:26Z explicitly identified R358's claim as wrong (memory 4351). Public arc memory 4349 describes the earlier Legendary unlocks.
• W38 story composer call 11710 nevertheless returned "their first Legendary ever"; persisted message 4198 and public story memory 4352 repeat it. Its captured prompt contains the Mother Witch/Phoenix arc context as well as the erroneous prior-post preview/spotlight summary. The error is therefore propagated prior output and unproved first-of-rarity semantics, not an absent recorder event.
• The weekly recap job posted normally and emailed 14, with receipt memory 4353. Successful delivery does not establish factual acceptance.
• Separately, R359 still says 30 straight winning weeks while the finalized week-close post 265 and W38 story say 31. Memory synthesis flagged this draft discrepancy too. Do not auto-approve or edit either card; leaders own their disposition.

Report continuity is present but one quantitative direction is false

• W38 member-report calls 11717-11730 all have prior-focus input and nonblank focus-review/next-focus blocks. All 14 eligible current recipients have receipts and persisted W38 focuses.
• Call 11717's prior-focus numbers are 2.78 elixir leaked in losses and 2.34 in wins, a 0.44 gap. Current brief says 2.43 in losses and 1.96 in wins, a 0.47 gap. The rendered focus review calls this a "tighter split than before." Both averages fell, but the quoted split widened by 0.03; do not conflate those statements.
• Call 11727 says "When you followed it" based on aggregate win/loss leakage, with no observed card-placement or timing evidence. Call 11720 compares a Ranked-specific prior intention to all-mode current leakage (209 battles, only 50 Ranked), then credits it for the Ranked climb. These are narrower comparability/attribution concerns, not evidence that members followed advice or that Elixir caused improvement.
• Current source computes only the current _elixir_profile; the previous window is reduced to _battle_tally in runtime/member_report.py. facts_for_model carries the prior intention as prose, not a structured measured prior/current comparison. The system prompt already says the intention is not proof of action, but parsing/rendering validates tag presence rather than quantitative direction or tactical follow-through.

Remediation and acceptance contract if approved

1. Establish canonical evidence for whether a newly observed card is the first of its rarity; unknown history must remain unknown. Reuse the existing collection/capability ownership rather than trusting a prior post summary. Cover genuine first unlock, pre-owned rarity, several simultaneous unlocks, and incomplete evidence. Preserve ordinary unlock narration and editorial discretion.
2. Supply deterministic prior/current comparisons from clan DB battle facts, with windows, mode scope, measured sample counts, and direction separated from tactical follow-through. Never make product behavior depend on admin-only telemetry or infer baseline numbers from advice prose. Cover the 0.44 -> 0.47 counterexample, mixed-mode comparisons, missing prior measurements, and insufficient samples.
3. Add proportionate regression/eval coverage for exact flagship output grounding. The current deterministic post-quality sample does not cover the email report narratives, and its clean score did not detect these failures; do not treat it as a complete scorecard.
4. With approval and clean synchronized preflight, acquire the agent lease only immediately before repository edits; prepare/stage only run-created paths; run focused regressions, relevant strict evals including Ask Elixir, and full gates; commit/push only current-run work. Restart only if runtime changes require it under the backed-up origin-owner exception.
5. Semantic acceptance is the next comparable natural unlock/story and next normal weekly report cycle: no unsupported first-of-rarity assertion, correct computed direction, explicit limits on comparability and unobserved tactics, normal delivery, and unchanged human/privacy boundaries. No synthetic request, event, post, repair, email, or early run. A missing comparable unlock through the following weekly cycle is insufficient_sample, not proof of success.

Current operational and supporting evidence

Clean synchronized main at 676e7cb; no objective lease. PID 45792 remains running. First normal post-restart tick 9170 completed on 2026-09-13T12:20:28Z; latest inspected tick 9376 has 47/47 members ready and no degraded materialization. Both live DBs pass quick_check. One-day quick confidence has zero findings; seven-day confidence retains four already-reconciled/recovered signatures. Seven-day strict Ask Elixir passes (17 messages), routing goldens pass 3/3, and deterministic post-quality samples 22 with zero flags. Strict leader actions fails on decision rate 0.846 and two stale proposed relays (R358/R359); trace and relay-copy coverage remain 100%. Human decision latency is not permission to settle those cards.

— Improve Elixir (agent)