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.

PIT-2 — the 2026-11-02 incident also cost a permanent point-in-time capture day (2026-10-30); the alarm fired, named the date, and nobody wrote it down — and every automatic witness to it goes blind on 2026-12-01

status: open · raised panel run 49 (2026-11-09), chair's own measurement · class: moat (NORTHSTAR asset #2 / gauge G4) + loudness-READER (G2) · judgement: YES · effort ~0.1 pd (two ledger entries) · horizon: hard deadline 2026-12-01


Plain-language summary for an owner reading one paragraph

On Monday 2026-11-02 the trading box failed its own test gate, so no morning run happened and GitHub's backup leg ran after the close instead. That incident is written up at length — the selection it produced was voided and the books were repaired (research/2026-11-03_momentum_selection_voided.md). But the missed morning run also cost something nobody has written down: the point-in-time snapshot for Friday 2026-10-30 was never taken and can never be taken again. That stream is the one the charter calls "the designed core of the survivorship-free forward record." The alarm for it worked perfectly — it fired that night, named the exact date, and failed the run. What failed was the reading: it fired into a channel the owner no longer reads by his own 2026-10-07 instruction, on the night of the biggest incident in two months, in a workflow leg that had concluded failure every day for the preceding ten days. It never fired again, because the check is one session deep by design. The shop did build a countdown for exactly this (expiring_gaps, 2026-10-19), and it works — but it only starts warning on 2026-11-23, with one week of calendar left. The day itself is gone. The record is still fixable, for about three more weeks.

Mechanism — measured and executed, not inferred

1. The alarm fired in production. GitHub Actions run 41560031438, job snapshot, 2026-11-03T01:34:25Z, verbatim:

  ✅ options_iv_term: 2026-11-02.parquet: 8126 rows, median 7 expiries/name
  ✅ shortability: 2026-11-02.parquet: 14397 names, 5305 ETB
  ❌ pit_snapshots: missing 2026-10-30, the previous session (the morning
trading run writes it): nothing was written for it
1 capture stream(s) FAILED integrity QA
##[error]Process completed with exit code 1.

That is PIT-1 leg 2 working exactly to its specification, on the first day it was ever needed. This row does not criticise the detector; it credits it.

2. Why that red carried no information. Three things converged on it:

  • config/fleet.yaml is notifications.email: weekly_digest_only (owner, 2026-10-07), so the failure was recorded, not mailed — by design. The steward reads those records in his place.
  • the leg it fired in had concluded failure on eleven consecutive runs: the Skew Snapshot backup leg reads failure on runs 116–126 (2026-10-19 → 2026-11-03), run 115 (10-16) being the last prior success and 127 the next. Runs 116–125 failed in 4–5 seconds each — too fast to have captured anything, a different and earlier defect (the daily audit's property, #321, "the backup legs can no longer reach the market"; cited here only as the noise floor). Run 126 is the exception and the one that matters: it ran 5m24s, did the captures, and carried the pit red below. So the one informative red of the month arrived as the eighth consecutive red in that leg — ambient, and indistinguishable from the others at a glance.
  • it fired on the night of the voided-selection incident, which consumed the next three days of attention.

The steward's 2026-11-05 report covers #318, #320, REF-1 and the shortability clock; the daily audit's 2026-11-05 report covers beacons and backup legs. Neither mentions it, and nor does anything else: grep -rln "missing 2026-10-30" --include=*.md over the whole repo returns nothing.

3. It never fired again — one session deep, by construction. Executed here:

_pit_missing_session(files, today)        expected    result
2026-11-02                               2026-10-30  "missing 2026-10-30 … nothing was written for it"
2026-11-03                               2026-11-02  None
2026-11-04                               2026-11-03  None
2026-11-05                               2026-11-04  None
2026-11-06                               2026-11-05  None

capture_qa('data', today=2026-11-05) -> pit_snapshots ok=True "2026-11-05.parquet: 885 rows"
capture_qa('data', today=2026-11-06) -> pit_snapshots ok=True "2026-11-05.parquet: 885 rows"

The check compares the store against pit_expected_date(today), a single date, so the red was available on session 11-02 alone. This is not a defect and not a test weakness — tests/test_data/test_pit_floor.py passes 19/19 here and its lag control correctly asserts the one-session-lag green; a deeper check would be GAP-1's residual 1, not this row. It is stated because it is the reason the record must now be written by hand: there is no second chance at the alarm.

4. The day was never captured, and the slot discipline worked for the day either side of it. git log --diff-filter=A -- data/pit_snapshots/2026-10-30.parquet is empty — never written, so this is a skipped slot, not a thin file:

fileadded by
2026-10-29.parquet54a5f4f — box run 2026-10-30 07:53 PT
(2026-10-30)nothing — no run existed on 2026-11-02 morning
2026-11-02.parquet372b1d8 — box run 2026-11-03 07:55 PT, 885 rows

The after-close 11-02 leg added no pit file, so the 11-02 slot stayed open and the next morning's complete run filled it honestly (885 rows, genuine 11-02 closes — A 171.99 against 10-29's 181.15, versus the 886-row norm CAPTURES.md pins). PIT-1 leg 1 saved 11-02. Then the 11-03 run's prices["date"].max() was 11-02, so it wrote 11-02 and never looked back; 10-30 had no later claimant. This is PIT-1's shape 2 (stale-date slot consumption), which PIT-1 shipped deliberately green for a stated reason, delegating the red to the check in §1/§3.

Only pit_snapshots lost the day, and only one day. order_log and portfolio_snapshots both hold 10-26…11-06 unbroken (checked). The 11-02 snapshot was separately tested for contamination by the 37-name partial fetch and is sound: on an 885-name join only 2 closes are identical to either neighbour, Spearman rank correlation is 0.9760 against 10-29 and 0.9953 against 11-03, and top-50 overlap is 44 and 48 respectively — re-measured directly in this run. So the moat loss of the 11-02 incident is exactly 2026-10-30 — no more, and no less.

5. It is in no ledger. '2026-10-30' in ACKNOWLEDGED_GAPS['pit_snapshots'] → False (the dict holds 07-15, 07-16, 08-05 and the eight September outage days). grep "2026-10-30" CAPTURES.md → no match, though § pit_snapshots records every prior notch in detail. The incident memo audits the run's other artifacts explicitly — "The equity row, the portfolio snapshot and the macro files from that run are real observations and stay as they are" (research/2026-11-03_momentum_selection_voided.md:50) — and never reaches the capture that the missing run failed to write.

6. The deadline is real, and a purpose-built countdown exists — which narrows this row's claim and sharpens its date at the same time. capture_gaps uses a 20-trading-day window ending at the stream's newest capture (captures.py, lookback_td: int = 20 — GAP-1 residual 1). Simulated against the real trading calendar:

newest capture 2026-11-25  window 2026-10-29 .. 2026-11-25   10-30 visible: True
newest capture 2026-11-27  window 2026-10-30 .. 2026-11-27   10-30 visible: True
newest capture 2026-11-30  window 2026-11-02 .. 2026-11-30   10-30 visible: False

The file dated 11-30 is written by the 2026-12-01 morning run, so 2026-10-30 leaves every automatic witness that day — the day the December go-live gate opens. Fifteen trading days from today.

Correcting an earlier draft of this row, which claimed thales captures' interior-gap line was "the last instrument": it is not. captures.expiring_gaps() (captures.py:491+, wired into thales captures) was built on 2026-10-19 for exactly this failure mode — its docstring says so, naming the 2026-09-21 day that "left every option stream's window ... No layer said anything, because by construction none could" — and CAPTURES.md:30 advertises it: "thales captures now counts each unacknowledged gap down to the session it becomes invisible … so the deadline announces itself." So the shop already built the fix for this class. Two measured facts about it:

  • It returns {} today. Executed: expiring_gaps('.', today=2026-11-09) → {}. Its warn_td is 5, and 2026-10-30 currently has 15 sessions left, so it is correctly silent — not broken.
  • It starts counting down on 2026-11-23. Computed from the calendar: 2026-10-30 enters the ≤5 band when the stream's newest capture is 2026-11-20, which is written by the 2026-11-23 morning run, then counts 5→1 and goes silent after 12-01.

So the honest statement is not "nothing will warn" but: the warning arrives with six routine slots left and one week of calendar. Routine slots inside the countdown: the panel on 2026-11-23 and 2026-11-30 (Mondays 14:00Z), and the steward and daily audit on 2026-11-26 (a Thursday — note it is not a trading day, Thanksgiving, so the routines fire on cron while no capture does). Today's own slot is three weeks earlier than any of them, which is the whole reason to file now rather than rely on the countdown.

Why this is worth a row at all — the honest accounting

The data cost is one day in ~110: near nil, and this stream already carries 11 entries in ACKNOWLEDGED_GAPS and about 15 recorded notches once CAPTURES.md's extra entries (2026-06-16, 2026-08-26 and the two truncated files 10-12/10-13, which it says to "treat as missing") are counted. The cost is to the ledger, which is the instrument behind charter gauge G4 ("capture-day gaps (target zero — each is permanent) … measured by thales captures / capture-qa, machine-checked"). An unledgered notch is precisely how that gauge goes quietly false — this repo has the worked example in GAP-1's 2026-06-17, a hole now ~100 trading days old, in no ledger, found only by archaeology. So the asset at risk is trust (asset #1), not research power.

Prior art — checked before filing

  • grep -ril over all four research/queue/ directories for pit_snapshots, PIT-1, interior gap, slot consumption, STAMP-2, notch; plus research/RESEARCH_QUEUE.md (frozen archive), TECH_DEBT.md, RESEARCH.md, AUDIT.md, CAPTURES.md, web/public/data/killlist.json. No pit-2 slug exists anywhere in the tree or the archive.
  • dismissed/gap-1-capture-day-census.md is the governing prior art and its own tripwire fires here: "Reopen immediately … if a seventh permanent capture-day is found", and "one ledger line in ACKNOWLEDGED_GAPS closes it". GAP-1 was dismissed as mechanical overflow (ranked 17 of 21 at the cap), not on merit. ops/RESEARCH_PANEL.md §4 does not sanction dismissed → open — which is why runs 28 and 32 annotated and escalated instead — so this is filed as a new slug for a new notch (GAP-1's named days are 2026-06-17 and 09-18→09-29, all outside the detector window), with an annotation added to GAP-1 so the two are never read as duplicates.
  • Said plainly, because it weakens this row's claim to novelty: GAP-1's condition had ALREADY been declared fired before today. Triage's 2026-11-03 dismissed sweep named it as item 3 of 11 fired conditions — "gap-1 — fired 2026-09-25, 39 days" — and handed it to the steward with "dismissed → open is the steward's move, never mine." And ops/STEWARD.md §4 does sanction that move — "reopen a dismissed item only when its own written reopen condition has fired — cite it" — so the authorised route exists and the actor already has GAP-1 on its desk. This row is therefore not the first firing of "a seventh permanent capture-day"; it is the first dated, still-visible, still-recordable instance of an already-fired condition. That is a weaker novelty claim and a stronger claim on the steward's attention, and it is the honest one.
  • built/pit-1-snapshot-green-while-empty.md — this is its shape 2, left green on purpose. Not a relitigation: legs 1 and 2 both shipped, both are credited above with production evidence, and no code change is proposed.
  • dismissed/cqa-2-latest-file-only-qa.md — PIT-1 carried its missing-today leg; condition (d) is satisfied, not reopened.
  • open/cqa-3-capture-qa-blind-to-capture-instant.md — a different blindness (capture instant, option streams). No overlap.
  • Not on the kill list; proposes no trial, so it costs no results/research_registry.jsonl row and no deflated-Sharpe budget.

Fix shape (propose-only) — the ledger paragraph is the load-bearing leg

Ordered deliberately, because CAPTURES.md:31-33 says which one matters: "Recording a notch in THIS file is what makes it permanent; ACKNOWLEDGED_GAPS only silences the daily alarm and must cite the record." An earlier draft of this row had the order the other way round.

  1. PRIMARY — one dated paragraph under CAPTURES.md § pit_snapshots, in the same form as the existing notch entries: 2026-10-30, D-1 of the 2026-11-02 morning run that never happened (box test gate, #316); the after-close backup leg stamped 11-02 and no run ever claimed 10-30; cite research/2026-11-03_momentum_selection_voided.md and Actions run 41560031438. This leg is the permanent one and survives 2026-12-01.
  2. SECONDARY — one entry in ACKNOWLEDGED_GAPS["pit_snapshots"] citing that record, which silences the daily alarm and keeps the date visible as a dim notch. On its own it is not enough, and after 12-01 it is cosmetic.
  3. Optional, one line, and worth more than this row — a standing doctrine sentence in CAPTURES.md § pit_snapshots: a missed morning run costs the previous session's snapshot, permanently; record it that night. The generalisation is sound and was checked: because pit is stamped D-1 while order_log and portfolio_snapshots are same-day, a missed morning run costs exactly one pit day and nothing else, by construction. One sentence retires this whole row class. It is prose, not a detector, so it is offered here rather than withheld.

Note for whoever builds it: this PR cannot auto-merge, so the steward's merge is a requirement and not merely a recommendation. ^src/ is an explicit DENY_PATTERNS entry in scripts/automerge_guard.py, so leg 2 is denied outright; and CAPTURES.md matches neither ALLOWED_PATTERNS nor DENY_PATTERNS, so leg 1 falls closed. Executed: automerge_guard.py --files CAPTURES.md,src/thales/data/captures.py → DENY: CAPTURES.md is not a record file (records-only fast path).

No code change is proposed beyond the one dict entry. Widening capture_qa beyond one session, or giving capture_gaps a full-history mode, is GAP-1 residual 1 and stays its property.

Test design / verification

  • After the entry lands, thales captures must stop printing the interior-gap warning for pit_snapshots while still listing 2026-10-30 through acknowledged_gaps_seen() — the ledger exists to explain holes, not hide them (captures.py docstring: "Reported so an acknowledged gap stays VISIBLE").
  • Negative control: remove the new entry and the warning must return, naming 2026-10-30. An entry that silences the warning without the date appearing in the acknowledged-gaps report is the wrong fix and must be rejected.
  • This verification expires with its subject, which is why leg 1 leads. acknowledged_gaps_seen() uses the same 20-trading-day window as capture_gaps (captures.py, window = [...][-lookback_td:]), so after 2026-12-01 both the warning and the dim-notch line go silent and neither the check nor its negative control can be produced any more. A motivated operator cannot wriggle out of the kill criterion, but it could wriggle out of this check by letting the date pass. Leg 1 — the CAPTURES.md paragraph — has no such expiry, and after 12-01 it is the only leg that can be verified at all (by reading the file).

Pre-registered kill criterion

Drop this row and record why if either holds:

  1. The day is reconstructible as an honest as-seen-live observation — i.e. thales snapshot-pit --date 2026-10-30 can rebuild it from owned data/raw and the ledger's doctrine permits it. The honest reading is that it cannot: § pit_snapshots' "Why time-gated" says vendor re-fetching re-bases adjusted prices, so "what the system actually saw on a date is irreproducible afterward", and the ledger already classifies the eight September days as a PERMANENT NOTCH on exactly this ground even though their slots were likewise empty. (PIT-1's 2026-09-28 annotation called those days "backfillable"; the ledger's permanent classification supersedes it. If the owner rules the other way, this row dies and GAP-1's September entries need the same correction — stated so the inconsistency cannot be resolved silently in this row's favour.)
  2. 2026-10-30 is already recorded somewhere authoritative this run did not read. Tested: grep -rln "missing 2026-10-30" --include=*.md over the repo returns nothing, and neither 11-05 report mentions it.

This row does not reopen if capture_qa stays one session deep. That is PIT-1's stated design and GAP-1's residual, not this row's claim.

Where this ranks, and who can close it

Not rank 1 on urgency — that claim was in an earlier draft and it is false. Under NORTHSTAR.md §4's own priority order, P0 (loudness & safety) outranks P3 (moat), and at least three open rows are P0 and live every session: dro-1 (every VRP open is a sell, so the one exposure alarm inverts), rwg-1 (the runaway-order guard, on the live path) and gtr-1 (a P0 detector that cannot fail on its own shape). Any of those beats a recorded notch.

Rank 1 on irreversible expiry, which is the claim that is true and still the reason to act: it is the only open row whose subject becomes invisible on a computed date, and fin-3's competing "must land before the December gate" horizon was mooted on 2026-11-05 when the gate-holder deferred it past December. On size it is the smallest row in the tree — one paragraph, one dict entry, ~0.1 pd.

It must not be routed to the implementer. The next implementer fire is 0 17 5 * * → 2026-12-05, and triage's is 2026-12-03 — both after the blindness date, and triage's 12-item cap means its 12-03 sweep could overflow-dismiss this row two days after the deadline it was filed to beat. ops/STEWARD.md §3 lists the trading path as src/thales/{execution,portfolio,strategy,features}, backtest/engine.py, data/downloader.py, data/macro.py, paper-trading*.yml, skew-snapshot.yml, scripts/win/ — src/thales/data/captures.py is not on it, so this is observability-tier work the steward may author and merge.

Recommended route: the steward authors and merges it on 2026-11-12 or 2026-11-19 — or it does not happen. open/ is not a notification channel, so this run also carries the item in its report and PR body rather than relying on the directory to deliver it.


BUILT 2026-11-12 by the steward (not the implementer), awaiting the 2026-11-19 steward's merge

Built along the route this row recommends: the implementer's next slot (12-05) falls after the 12-01 blindness date, and src/thales/data/captures.py is not on STEWARD.md §3's trading-path list, so this is observability-tier work the steward may author. It must not be merged by the run that wrote it (§0), so the 2026-11-19 steward reviews it cold and merges it. That still lands before 12-01, with one more slot (11-26) as a fallback.

  • Leg 1 (the record): dated paragraph "SIXTH NOTCH — 2026-10-30" under CAPTURES.md § pit_snapshots, citing #316, Actions run 41560031438 and research/2026-11-03_momentum_selection_voided.md.
  • Leg 2 (alarm silencer): ACKNOWLEDGED_GAPS["pit_snapshots"]["2026-10-30"], whose text cites the CAPTURES.md record.
  • Leg 3 (the standing rule): one sentence in the same section: a missed morning run costs the previous session's snapshot permanently; record it that night.
  • Verification, run against src/: thales captures no longer prints the interior-gap warning, and lists pit_snapshots 2026-10-30 — D-1 of the 11-02 morning run… as an acknowledged gap. Negative control: with the dict entry removed, the warning returns, naming 2026-10-30. The new test test_pit_2026_10_30_notch_is_recorded_not_just_silenced pins both legs and fails without the entry.
  • Kill criterion: neither condition holds. The ledger's doctrine treats a missed slot as not reconstructible (the September days are classed the same way), and no record of the date existed before this one.