Thales
← research journal

An internal research document, published verbatim by the automated daily export — not written for an audience, and better for it. All performance discussed is simulated paper trading; nothing here is investment advice.

GCA-1 — DISMISSED ON ARRIVAL: "the December gate's day count disagrees across three records" — the arithmetic does not reproduce, and the proposed fix would have put the error INTO the record it was protecting

status: dismissed · raised AND killed in panel run 50 (2026-11-16) · class: would-be P1 registration · never filed as open


What was claimed. That thales go-live-gate's 108 live trading days projected forward gave 118 on 2026-12-01 · 125 on 12-10 · 126 on 12-11, and that therefore AUDIT.md:128's fire-once register row ("reaches 119 on 2026-12-01 and 126 on ~2026-12-10") was off by one, as was research/steward/2026-11-12.md's carried-forward line.

Why it is wrong. The count's anchor was misread. thales go-live-gate prints "108 … as of 2026-11-16", but data/processed/equity_history.jsonl's last row is 2026-11-13 — today's box run fires at 15:35Z and has written nothing. So 108 counts through 11-13, and 2026-11-16 itself is an uncounted trading day. Both the queue-and-clocks lens and the panel chair walked forward from 11-17 and dropped it.

Counting 11-16 onward inclusive (thales.utils.calendar.is_trading_day; 2026-11-26 Thanksgiving confirmed a non-trading day):

daten_live
2026-11-16109
2026-12-01119
2026-12-10126 ← first day ≥ 126
2026-12-11127

AUDIT.md:128 is CORRECT. Independently corroborated: AUDIT.md:247's 2026-11-01 weekly-audit entry derives the same 119 / 12-10 from a 98-day read, and 98 + the ten trading days 11-02…11-13 = 108, so no days were lost and two derivations three weeks apart agree.

Why this kill is worth recording rather than quietly dropping. The register row exists precisely so that "a NOT-PASS driven purely by min_live_trading_days is never quoted as a forward-performance result." The proposed remedy would have written a one-day error into that row — a finding whose fix degrades its own target. It was also a correlated error: the lens made it and the chair reproduced it with the same loop before catching it, and the adversarial reviewer derived the correction independently.

The one true residue, handled as a carry-forward, not a row. research/steward/2026-11-12.md:35 reads "12-10: the earliest date the December gate can be decided (107/126 that day)". The 107 is that run's own as-of count mislabelled as 12-10's; the correct figure is 126/126. The sentence's headline ("12-10 is the earliest date") is right and agrees with AUDIT.md, so the parenthetical contradicts its own clause and reads as a typo. Carried to the 2026-11-19 steward run in panel run 50's report and beacon marker.

Reopen conditions (any one)

  1. The gate's day count is quoted in a decision record as a forward-performance result without its day count alongside it. That is AUDIT.md:128's own pass criterion, still ⏳ pending, and it is the thing this candidate was reaching for.
  2. go-live-gate's min_live_trading_days basis changes (marks vs returns, or the clean-clock since moves) without the register row being re-derived — at which point three records would genuinely diverge rather than one carrying a typo.
  3. A trading day is actually lost (an outage with no equity row) between now and 2026-12-10, which pushes the 126th day later than the register row projects. The row already caveats this ("assuming no missed days — any outage pushes it later"), so the reopen is for the case where the caveat fires and nobody updates the date.