CQA-3 — no failing check reads captured_at, so two whole skew-capture days taken after the close passed every gate — and the rule that would disqualify them has been awaiting your countersign since the day the field was born
status: open · raised panel run 25 (2026-09-17; frontier-WATCH catch, verified independently from the committed store — a defect row, not a research or capture proposal, so it does NOT discharge the scout's 8-week cadence clock; same treatment as BTB-1) · class: moat continuity (G4) + registration hygiene (G5) · effort ~0.25 pd · horizon: live daily, and it recurs on every double-failure session · PRECONDITION: not approvable for build until you rule on the 2026-09-02 B4 append — see "The ruling this forces"
Plain-language summary for an owner reading one paragraph. On 2026-09-14
and 2026-09-15 the option-skew capture — the crown-jewel forward stream — ran
entirely after the market closed: every one of the 889 and 888 rows was
taken at 18:24 ET and 18:03 ET respectively, when quotes are stale and wide.
Both days passed every quality gate with a perfect score, because no failing
check anywhere reads the captured_at field, and the row counts looked
normal. They are not recorded as displaced in CAPTURES.md either. A draft
rule that would disqualify exactly these days was written into the B4
activation pin on 2026-09-02 — the same day the captured_at field was
introduced — but it is marked "binds only with the owner's countersign" and
you have not ruled on it. So these two days are the first real data that force
that ruling, and we found out about them by hand, not from an instrument.
Mechanism — verified at HEAD dfde575
1. The two days are post-close, and only captured_at says so. Read
directly from the committed data/options_skew.jsonl with stdlib json (the
package does not install in this container — see "What did not run"):
| date | rows | distinct captured_at | ET | healed |
|---|---|---|---|---|
| 2026-09-11 | 889 | 19:11Z + 21:38Z | in-session + top-up | unset |
| 2026-09-14 | 889 | 22:24:02Z only | 18:24 ET — post-close | unset |
| 2026-09-15 | 888 | 22:03:07Z only | 18:03 ET — post-close | unset |
| 2026-09-16 | 884 | 19:16:04Z + 22:01:07Z | in-session + post-close top-up | unset |
Nothing else distinguishes them. healed is unset on all four days, so the
mixed-provenance marker does not fire either.
2. No failing check reads the field. grep -n "captured_at" src/thales/data/captures.py returns zero hits — the QA module never reads
it at all. The options_skew branch of capture_qa
(src/thales/data/captures.py:712-790) tests only per-date row counts against
a trailing median, plus unparseable lines. 889 and 888 against a trailing
median of ~889 sail through. Across the whole src/ tree the field is read in
exactly two places — data/options_skew.py (which writes it) and
data/resession.py (the re-session tool, which reads it per row). So the
honest claim is narrow and is the one that matters: the field is
machine-readable, and no check that can go red looks at it.
3. The manual register was not updated either. CAPTURES.md records
displaced days for 08-06/07 and 08-27/28/29 only. Nothing records 09-14 or
09-15. That matters because the draft rule's own clause (b) makes
CAPTURES.md the register of record for days whose basis is in question.
4. The writer already owns the instrument it does not use.
session_phase() exists in src/thales/utils/calendar.py, and
snapshot-shortability already uses it to REFUSE a pre-open capture (STAMP-3,
built). snapshot-skew (src/thales/cli.py:4656-4664) calls
market_session_date() and is_trading_day() and never checks the phase.
That asymmetry between two captures in the same repo is the sharpest form of
this finding: one stream knows what time it is, the other does not.
5. How the two days happened, for context only. Both the launchd dispatch
agent (ops/launchd/com.thales.dispatch-skew.plist:15 — "Mon-Fri 12:00 LOCAL
= 15:00 ET — in-session, an hour before the close, the capture's intended
time") and the backup cron in skew-snapshot.yml (19:00Z = 15:00 ET) are
configured in-session. Landing post-close needs BOTH to fail: the agent
produced no run on 09-14 or 09-15 (single skew commit each day, 22:29Z and
22:10Z; the Actions record shows no workflow_dispatch for skew-snapshot.yml
on either date) and GitHub's cron then arrived ~3 h late. That conjunction
is supposed to be rare. It happened on 2 of the 9 days on which
captured_at exists at all (the field landed 2026-09-02), consecutively — a
22% base rate over the only window in which the question is measurable. The
ops-side half of this was considered separately this run and DISMISSED
(dismissed/dsp-1-dispatch-liveness-attested-on-send.md): every session in
the window still traded in-session on the backup cron, so there is no trading
harm to claim. This row is about the data and the gate, nothing else.
The ruling this forces — why this row is not approvable yet
research/2026-08-01_b4_skew_activation_pin.md:235-258 is headed
"2026-09-02 append — PROPOSED refinement (binds only with the owner's
countersign)" and closes "if the owner declines it, the field still exists
and (a)–(c) simply do not apply." Its three clauses:
- (a) a signal date's rows count as valid only if
captured_atfalls inside the NYSE session (09:30–16:00 ET); - (b) rows without the field are assumed in-session unless
CAPTURES.mdrecords the day as displaced; - (c) a date whose only capture is post-session does not qualify as a signal date.
No countersign exists — not in the memo, not in CAPTURES.md, nowhere in
research/. The append is 15 days old and has been pending for the entire
lifetime of the field it governs. CAPTURES.md's own B4 amendment summary
(:136-160) does not mention (a)–(c) either, so a reader consulting the
ledger rather than the memo would not know the question is open.
Note also that the neighbouring stream has already ACCEPTED this condition:
CAPTURES.md:281-287, inside ### options_iv_term, carries a standing
"Post-close quote caveat" telling consumers to treat a day's row as an
approximate EOD mark. So the two streams, produced by the same workflow run,
currently sit under different and unreconciled basis standards. That is
the registration-hygiene half of this row.
So the decision tree is:
- You decline (a)–(c) → there is no rule to enforce; the row collapses to
a single
CAPTURES.mdline recording 09-14/09-15 as post-close, plus (if you want it) a reconciliation of the skew stream's basis standard withoptions_iv_term's accepted caveat. A record, not a build. Triage must not approve the build ahead of this ruling. - You countersign (a)–(c) → the two days become non-qualifying signal dates and the gate below is what makes the rule real instead of aspirational.
What to build, if the rule is countersigned
A single failing condition in capture_qa: a capture date whose rows ALL
carry a captured_at outside 09:30–16:00 ET fails, naming the stream and the
date, and such days get written to CAPTURES.md's displaced-day register
per clause (b). ~0.25 pd. The cheaper half — recording 09-14/09-15 in
CAPTURES.md — is worth doing under either branch of the ruling, because the
raw captured_at values are the only record today and clause (b) points a
future reader at a register that does not mention them.
Test design and the REQUIRED negative control
The control is free and comes from real data, and it is what keeps the gate from being over-broad:
- must RED on 2026-09-14 (every row post-close), and
- must stay GREEN on 2026-09-16, which carries an in-session
19:16:04Zcapture and a post-close22:01:07Ztop-up of 5 names.
A gate that reds 09-16 is wrong and must not ship: mixed provenance is the
normal, healthy refire pattern that captured_at/healed exist to express,
and SKW-3 (built) already fixed the real hazard in that shape. This row is the
next axis of SKW-4 leg 1 (built), which filtered rows stamped on a
non-session date; this one is about rows captured outside the session
clock. Cite both so triage does not re-derive them.
Kill criterion — pre-registered
- Dies mechanically at pickup if
grep -rn "captured_at" src/thales/data/captures.pyis non-empty and a fixture date whose every row carries a post-16:00-ETcaptured_atreturnsok: Falsefromcapture_qa. One command, one run. - Dies on the ruling if the owner declines the B4 append — see above.
- Premise falsifiers, both already checked and negative:
captured_aton 09-14/09-15 falls inside the session (it is 18:24:02 and 18:03:07 ET); orCAPTURES.mdor an existing memo already records those days as displaced (neither does).
What is NOT claimed
- No statistical damage has occurred. B4's primary is a weekly cross-sectional rank-IC with a ≥400-valid-row floor. 2026-09-16 is in-session and well over the floor, so ISO week 38 keeps a qualifying signal date under clause (c). The defect is that the condition was invisible and found by hand, not that a number is wrong.
- No trading harm. Every session in the window traded in-session on the backup cron.
- This is not CQA-2. CQA-2 is
files[-1]-only reads (a missing today-file failing nothing) plus alert-subject attribution; its missing-today check would have PASSED on both days, since the files exist with normal row counts. Same module, orthogonal defects — fold partners for triage, neither subsuming the other.
Effort · horizon
~0.25 pd for the gate plus its two-sided control; the CAPTURES.md record is
minutes. Live daily. research/queue/open/ is at its 12-item cap, so this
evicts an incumbent — the ranking argument, no more: a quarter of a
person-day, damage that is permanent per occurrence, a 2-in-3-session
observed rate, and a pending owner ruling that this data forces either way.
What did not run
pip install . -c requirements-ci.lock --no-build-isolation FAILS in the
routine container (ModuleNotFoundError: No module named 'hatchling'), so
thales, polars and pytest are unavailable. Every number above was read
from the committed store with stdlib json, or from git log, or from the
GitHub Actions API via the session's GitHub tools. Nothing was run against a
broker, and no capture, config or trading path was touched.
TRIAGE ANNOTATION 2026-09-21 — base rate measured; the precondition is now 19 days old
No recurrence since the two days this row is about. Re-read from the
committed data/options_skew.jsonl: 09-16 19:16:04Z + 22:01:07Z, 09-17
19:01:43Z + 22:01:40Z, 09-18 19:09:37Z + 21:34:28Z — healthy mixed
in-session/after provenance, which the row's own required negative control says
must stay green. Base rate stands at 2 post-close-only days of 12 stamped
days. The mechanism is unchanged: grep -c captured_at src/thales/data/captures.py = 0; cli.py:4655-4664 still checks only
market_session_date() + is_trading_day(), never session_phase().
The precondition is the binding constraint, not the build. The 2026-09-02
B4 append (research/2026-08-01_b4_skew_activation_pin.md:233-258) is still
uncountersigned — grep -rn countersign research/*.md CAPTURES.md finds no
ruling on clauses (a)-(c) anywhere. 19 days. Until it lands this row is not
approvable, by its own header.
Fold, per CQA-2's overflow dismissal this run: when this row is built,
carry CQA-2's subject-attribution leg (name the failing stream in the 🔴
subject) into the same change — both live in captures.py and the alert path,
and CQA-2's reopen condition (c) fires if they are built without it.
Triage annotation — 2026-09-28 (annotation only; the body above is untouched)
Still live at HEAD 864da4e. grep -c captured_at src/thales/data/captures.py
= 0. snapshot-skew still calls only market_session_date() +
is_trading_day(); session_phase is imported at cli.py:12 and used at
exactly one site, cli.py:2201 (the shortability writer). capture_qa's
options_skew branch (captures.py:714-790) is row-count and parse only.
CAPTURES.md:39-69 still records displaced days for 08-06/07 and 08-27/28/29
and nothing for 09-14/09-15.
Numbers unchanged, and unchanged for a reason. From data/options_skew.jsonl:
71 stamped dates, 2026-06-09 → 2026-09-18. The two post-close-only days are
still 09-14 at 22:24:02Z (889 rows) and 09-15 at 22:03:07Z (888); 09-16,
09-17 and 09-18 all carry healthy mixed in-session + post-close provenance. Base
rate 2 of 12 stamped days, identical to the 09-21 reading because no
capture has run since 09-18 (GitHub Actions refused every job repo-wide from
09-21; allowance resets 2026-10-01). The two-sided control the row needs — RED
on 09-14, GREEN on 09-16 — is still available in committed data.
THE MATERIAL CHANGE: the pending ruling stopped being a records question and
became an operational one, inside four days. Panel run 29 (2026-09-28, PR
#188) established a hard zero-margin deadline that this row now collides with
head-on. The hash-frozen B4 pin
(research/2026-08-01_b4_skew_activation_pin.md:83-85) drops an observation
when the gap to the next signal date exceeds 10 trading days. Last capture
2026-09-18; NYSE 09-18 → 2026-10-02 is exactly 10 (kept), 10-05 is 11
(destroyed, permanently). If Actions is still dark the remedy is a hand-run
thales snapshot-skew from the Mac — and a hand-run after the close is
precisely the post-session capture that clause (c) of the uncountersigned
2026-09-02 append would disqualify. So the rescue action and the pending rule
are in direct conflict: under (c) a post-close rescue saves nothing, and under
no-(c) it saves the observation. Nobody knows which, because the append is
still uncountersigned at day 26 (:234-236, "binds only with the owner's
countersign"; the only countersign token in the file is that header, and no
ruling exists anywhere in research/*.md or CAPTURES.md).
Consequence for ranking: the ruling must be made BEFORE the rescue capture, not after. Deciding what counts as a valid row after seeing whether the rescue landed in-session is the after-the-fact reinterpretation this shop treats as a cardinal sin.
Blocked on: the owner's (a)-(c) ruling — unchanged, and by this row's own
header it is unapprovable for build until that lands. Effort once ruled: ~0.25
pd (one predicate + one fixture pair). Recording 09-14/09-15 in CAPTURES.md
is minutes and is worth doing under either branch of the ruling.
PANEL RUN 32 ANNOTATION 2026-10-01 — the countersign is FREE ON THE ACCRUED SAMPLE, and today is the last clean moment to make it
Why this annotation exists. This row's own header makes it unapprovable until the owner rules on the 2026-09-02 B4 append. That ruling has been pending 29 days, and nobody had measured what adopting it would cost. This run measured it. The answer is: nothing, on everything accrued so far — which removes the only substantive reason to keep deferring.
What was measured
Read from the committed data/options_skew.jsonl with the pin's own rules
(signal date = LAST capture date in an ISO week carrying ≥400 valid rows;
observation dropped when consecutive signal dates are >10 trading days apart),
applying clauses (a)+(b)+(c). Clause (b)'s displaced days are taken
post-ruling — {2026-08-06, 2026-08-27, 2026-08-28}, i.e. where the
2026-09-06 owner re-session moved the mis-stamped rows to; 08-07 retains only
its 5 genuinely in-session rows and is under the floor regardless.
Five dates flip from qualifying to non-qualifying:
| date | rows | in-session rows | why |
|---|---|---|---|
| 2026-08-06 | 891 | 0 | pre-field, displaced (clause b) |
| 2026-08-27 | 894 | 0 | pre-field, displaced (clause b) |
| 2026-08-28 | 893 | 0 | pre-field, displaced (clause b) |
| 2026-09-14 | 889 | 0 | captured 18:24 ET — post-close (clause a) |
| 2026-09-15 | 888 | 0 | captured 18:03 ET — post-close (clause a) |
And zero observations change. Two ISO weeks do lose their last qualifying
date and fall back to an earlier one — W32 2026-08-06 → 2026-08-05 and W35
2026-08-28 → 2026-08-26 — but both resulting bridges land at 7 trading
days, four short of the drop threshold:
BASELINE (append declined) : 14 signal dates, 13 of 13 observations kept
WITH clauses (a)(b)(c) : 14 signal dates, 13 of 13 observations kept
2026-08-05 -> 2026-08-14 7 td KEEP
2026-08-26 -> 2026-09-04 7 td KEEP
The deadline anchor survives the countersign too. Clause (a) is a row-level test, so the mixed-provenance days lose rows as well as the two post-close days losing all of them. 2026-09-18 — the last qualifying capture and the anchor of the live 10-trading-day deadline — goes 888 → 885 in-session rows, far above the ≥400 floor. 09-16 → 879, 09-17 → 886, 09-11 → 888, 09-04 → 889. Adopting the append does not move the deadline.
What this does and does not license
- Free on the past. Adopting (a)(b)(c) costs zero accrued observations. Nothing already banked is lost, so the ruling cannot be accused of having been shopped against the accrued sample.
- NOT free forward, and the interaction is live tonight. The capture stream has been dark since 2026-09-18 (Actions allowance exhausted; runners returned 2026-10-01 01:18Z). A rescue capture must land by Fri 2026-10-02. If that rescue lands post-close and (c) is in force, it does not qualify, ISO week 40 gets no signal date, and the loss happens anyway despite a capture having run. The two post-close days this row is about (09-14, 09-15) arose from exactly today's regime: the Mac dispatch produced no run and GitHub's cron arrived ~3 h late.
- This is why the ordering matters. Tonight's capture instant is not yet known. Ruling today is a decision made blind; ruling tomorrow is a decision made after seeing whether the capture landed at 15:00 ET or 18:30 ET, on the single most valuable pending observation in the shop. Today is the last clean moment.
Scope of this annotation
Measurement only — no clause, threshold, sample or decision rule is proposed
or changed here, and the row's build, precondition, kill criterion and fold
instruction are untouched. No skew-vs-return statistic was computed (and could
not have been: data/raw/, the pin's §3 return source, is empty in this
container). This extends what the row already establishes for W38 — "no damage
in W38" — to the strictly stronger "no damage anywhere in the accrued sample".
ANNOTATION 2026-10-05 — panel run 34: run 33 told the owner a box/GitHub skew collision costs "a permanent capture day". That is WRONG in both directions, and the correction is recorded here because the wrong version was handed over as fact. Separately, this row's own urgency is discharged.
No row, no threshold moved, no new proposal. Two records only.
(1) The skew-collision harm statement is corrected
Context: from the tbx-win cutover the box runs the skew capture at 12:00 PT,
which in the current PDT regime is 19:00Z — the identical instant as
.github/workflows/skew-snapshot.yml:12's cron: '0 19 * * 1-5'. Run 33
escalated this to the owner as the run's "only unrepairable harm", priced as a
permanent capture-day loss. It is not that.
What a same-instant double-write actually does. snapshot_skew builds its
done set from its own working copy (src/thales/data/options_skew.py:429-438)
and skips any (date, symbol) already present (:486). That per-symbol
idempotent skip is the capture's only already-ran gate — the structural
equivalent of the trading legs' already_ran_today — and it is precisely the
thing that cannot engage at an identical instant, because neither writer's
checkout holds the other's rows. So both pay the full ~1000-name fetch and both
commit. A rebase then cannot merge them: reproduced in throwaway repositories,
CONFLICT (add/add) on the per-date parquets (both-added, whole-file
os.replace writes at :592-594 and :623-625) and CONFLICT (content) on
data/options_skew.jsonl (two appends to the same end-of-file hunk), rebase
exit 1.
The winner's data therefore lands COMPLETE. No capture day is lost, and
duplicate (date, symbol) rows cannot arise — the append conflicts loudly
rather than silently concatenating. Both halves of run 33's statement were
wrong, and in opposite directions: the irreplaceable asset is safer than the
record said, and what is actually exposed is a billed Actions capture, a red run
with its alert email, and — when the box is the loser — a leftover
.git/rebase-merge whose knock-on cost lands on the trading runner, not the
capture.
And the exposure itself is far smaller than "18 identical-instant weekdays"
suggests. Measured from the Actions API this run: the last 17 of 17
skew-snapshot.yml schedule deliveries arrived 150–286 minutes after the
19:00Z cron (earliest ever observed 21:30:57Z, 2026-09-10), while the capture
itself completes in 6.2–8.6 minutes of wall clock and the box starts at
19:00Z on the dot. A collision requires GitHub to deliver a schedule event about
two and a half hours earlier than it ever has in the measured record — and
PR #191's thesis is that the delivery regime gets worse, not better. "All 18
weekdays identical-instant" describes the configured coincidence, not the
realised exposure. The benign branch is the overwhelmingly likely one: GitHub
checks out after the box pushed, done is complete, zero API calls,
written=0, nothing staged, no commit, green.
Two corrections to this panel's own intermediate derivation, so run 35 does not inherit them either:
- The 2026-11-02 DST change raises the overlap, it does not reduce it. The box moves to 12:00 PST = 20:00Z while the cron stays 19:00Z, which moves the box's window toward the observed 21:30–23:46Z arrival band, not away from it. Neither regime is demonstrated; the direction stated earlier was backwards.
- The trading-state orphan chain (a leftover rebase poisoning the next task's
push and the sync after that orphaning its commit) is directional, not
established:
scripts/win/boot.py:91-94aborts a stale rebase first, before the fetch and park, which blunts the first hop, and the rest depends ongit checkout -f -Bsucceeding with.git/rebase-mergepresent — not verifiable from a cloud container.
No remedy is proposed, because the repo already pre-registered one and it is
gated on exactly the condition that has not occurred:
ops/windows-host/README.md:283-286 — "If GitHub ever delivers it on time, it
can collide … then move that cron, not the box's task (the 15:00 ET capture
time is part of the research record)." Nothing pins the cron string
(tests/test_execution/test_ops_alerting.py:839-849 asserts only that a
schedule trigger exists and that the concurrency group is set, so deleting
the schedule would red but moving it would not), and the DST-robust
non-colliding slot would have to clear 20:00Z plus the run, i.e. 0 21 * * 1-5
or later — which is at or after the close, so a genuine-backup capture would
carry drifted quotes. Degraded but self-describing via captured_at.
Reopens, sharply — these are the numbers the README's own trigger lacked:
- (i) a
skew-snapshot.ymlschedulerun is ever created within 90 minutes of 19:00Z; or - (ii) any date in
data/options_skew.jsonlshows two distinctcaptured_atclusters; or - (iii)
.git/rebase-mergeis ever observed on the box.
Any one of those converts the configured coincidence into a live one, and the pre-registered remedy applies without further deliberation.
(2) This row's own urgency is discharged — ruling today is neither blind nor contaminated
Run 32's annotation argued that 2026-10-01 was "the last clean moment" to
countersign clause (c) blind, because a rescue capture might still land and
become a signal date. Measured from the committed store this run: 2026-10-01
carries 281 rows (captured 19:05:49Z) and 2026-10-02 carries 282 rows
(19:01:11Z). Both sit below the ≥400 valid-row floor
(research/2026-08-01_b4_skew_activation_pin.md:71), so neither is a signal
date regardless of how clause (c) is ruled — and both landed in-session
(15:05 and 15:01 ET), so the capture instant was never the binding constraint.
The 10-02 deadline expired unmet exactly as dismissed/b4-notch-skew-row-floor.md's
two appends forecast; today's only addition is that the forecast is now
confirmed by the data rather than by arithmetic. Not re-ruled here.
Consequence for this row: the countersign can now be made with the sample fully in view and nothing at stake in either direction. The B4 2026-09-02 append has been uncountersigned for 33 days, and this row remains unapprovable by its own header until it is ruled.
Triage annotation — 2026-10-05 (annotation only; the body above is untouched)
Mechanism unchanged at HEAD 502690b: grep -c captured_at src/thales/data/captures.py = 0; src/thales/cli.py:4658-4664 still calls
market_session_date() + is_trading_day(session) and never session_phase().
The field is read only in cli.py (1 site), data/options_skew.py (8) and
data/resession.py (22) — nowhere in the QA path.
The run-32/34 finding is CONFIRMED on the data, and the window it opens is
the reason this row is in the decide list. Measured from
data/options_skew.jsonl: no post-close-only day has recurred since 09-18.
The two post-outage rescue captures are 2026-10-01, 281 rows @ 19:05:49Z
(15:05 ET, in-session) and 2026-10-02, 282 rows @ 19:01:11Z (15:01 ET,
in-session). Both fall below the ≥400-valid-row floor
(research/2026-08-01_b4_skew_activation_pin.md:71), so neither is a signal
date under any ruling of clause (c), and the capture instant was never binding
on them. The countersign is therefore free in both directions right now —
it costs nothing on everything accrued, and it cannot be accused of being
chosen to save or sink a specific observation. That is the cheapest it will
ever be, and the window closes the next time a post-close capture lands on a
day that clears the row floor.
Uncountersigned for 33 days. …b4_skew_activation_pin.md:235-236 reads
"binds only with the owner's countersign"; grep -n countersign on that file
returns line 236 and nothing else, and no ruling exists anywhere in research/
or CAPTURES.md. The row remains unapprovable by its own header until ruled.
Number corrected: the post-close base rate is 2 of 14 stamped days, not 2 of 12 — the field landed 2026-09-02 and the stamped dates are 09-03…09-18 plus 10-01 and 10-02.
A LARGER THREAT TO B4 THAN THIS ROW, found while verifying it, and recorded here because no row owns it. Captures of ~281-283 rows are a recurring class, not an outage artifact: 2026-06-29, 06-30, 07-01, 07-02, 07-06, 08-03, 08-31, 10-01 and 10-02 are all under 400. If ~281 is the post-cutover steady state rather than a bad-day shape, every future ISO week fails the ≥400 floor and B4's signal-date supply stops regardless of how the capture instant is ruled. That is a question about the capture's row yield, not its timestamp, and it belongs to the panel — flagged to it in this week's report rather than folded in here, because writing it into this row would change the row's substance.
Rank this week: 8 of 12 (P1 — the ruling is dated by opportunity, not by harm; the row's operational urgency is genuinely discharged for now). BLOCKED on: the owner's countersign, by the row's own precondition.
Panel run 35 append — 2026-10-06 (evidence only; the body above is untouched)
A THIRD ALL-POST-CLOSE DAY LANDED YESTERDAY, AND THIS TIME THE GATE WAS RUN
RATHER THAN READ. data/options_skew.jsonl's 282 rows for 2026-10-05
carry a single captured_at = 2026-10-06T00:43:01+00:00 — 20:43 ET, 4h43m
after the 16:00 ET close, the latest instant of any such day (09-14 was
+2h24m, 09-15 +2h03m). The options_chains and options_iv_term twins for
that date carry the identical instant. Delivery was the schedule backup leg,
run created 2026-10-06T00:42:09Z — 5h42m after its 19:00Z cron (GitHub
Actions API). market_session_date() stamped the session correctly, so STAMP-1
held; nothing else about the day is unusual.
The negative control this row previously could not run, now executed. Prior
appends established "no failing check reads captured_at" by reading
src/thales/data/captures.py. This run installed the package (see the run
report's setup note) and ran the gate against the committed store:
thales capture-qa→ every stream OK,options_skewincluded, with the 282-row shortfall attributedCALENDAR-EXPLAINED.thales captures→options_skewstale 0 (td), i.e. current.CAPTURES.mdcarries no post-close marking for 2026-10-05.
So the defect is now demonstrated by execution, not inference: a capture taken 4h43m after the close is indistinguishable, to every instrument the shop owns, from one taken at the intended in-market instant.
THE BASE RATE IS 15 OF 37, NOT 2 OF 14 — AND THAT CHANGES WHAT THE
COUNTERSIGN IS. This row has only ever been able to see days where
options_skew.jsonl carries the field (it is null for every row before
2026-09-03), which is 14 days. The options_iv_term twin has carried
captured_at since the stream's first day, 2026-08-03, and the two streams
are written by one run at one instant. Measured across all 37 term files
against each session's true NYSE close (pandas_market_calendars):
| the day's first capture instant | days | of 37 |
|---|---|---|
| in-session | 20 | 54% |
| mixed (in-session + a later leg) | 2 | 5% |
| entirely after the close | 15 | 41% |
The 15, with minutes past that session's close: 08-03 (+36), 08-04 (+39), 08-05 (+36), 08-06 (+284), 08-11 (+2), 08-12 (+3), 08-26 (+105), 08-27 (+452), 08-28 (+362), 08-31 (+190), 09-01 (+101), 09-02 (+102, and a second leg at +383), 09-14 (+144), 09-15 (+123), 10-05 (+283).
So the draft clause's bite depends entirely on where its threshold sits, and the bite is now measured at each candidate:
| disqualify a day whose first capture is past the close by | days disqualified of 37 |
|---|---|
| any amount (> 0 min) | 15 |
| > 30 min | 13 |
| > 60 min | 10 |
That is the decision-relevant change. The row has been asking for a yes/no on a clause whose cost looked like two anecdotes; the honest ask is a threshold on a condition that has held on roughly two days in five since the stream was born. A ruling that disqualifies "post-close" without a threshold retires 41% of the skew family's life, including six days nobody has ever flagged.
Already recorded vs genuinely new. CAPTURES.md:33-100 records the
date-stamp collisions and the displaced-cron regime that cover 08-06, 08-27,
08-28 and the through-08-31 window; CAPTURES.md:281-287 already concedes, as
a standing caveat on this very stream, that the cron "observedly fires ~2h late
… often at/after the 20:00Z close" and tells consumers to treat each row as an
approximate EOD mark. The genuinely unflagged all-post-close days above an hour
are therefore 2026-08-26 (+105), 09-01 (+101) and 09-02 (+102/+383) — plus
the two this row owns (09-14, 09-15) and yesterday's 10-05. The five small ones
(+2 to +39 min) are covered by that standing caveat as a class.
The witness, and its limit, stated honestly. The twin establishes the
session's first capture instant, not the full set of legs. On the 14 dates
where both streams carry the field the earliest instant agrees every time, but
the full leg SETS differ on 5 of them, in both directions: 09-11, 09-16, 09-17
and 09-18 have a later top-up in options_skew that the overwritten term
parquet does not hold, and 09-03 has one in the term parquet that the skew
stream does not. For the 23 pre-09-03 dates the twin is the only witness
available; it is sound for the first-instant claim and silent on whether an
additional leg existed. A day read as all-post-close could in principle have
had an earlier skew-only leg; nothing in the writer's resume-dedupe suggests it,
and the direction of the other observed divergences is later, not earlier.
Nothing here changes the row's conclusion, its kill criterion, or its precondition — it remains unapprovable until the 2026-09-02 B4 append is ruled, and the triage's 2026-10-05 finding that the countersign is "free in both directions right now" still holds (10-05's 282 rows are under the ≥400 floor, so it is not a signal date either). What changed is the quantity the ruling turns on. Recorded by panel run 35; no new row, no queue slot, no trial.
The ruling landed — 2026-10-15 (annotation only; the body above is untouched)
The owner countersigned the B4 pin's 2026-09-02 append (a)–(c) with a
60-minute grace past the close. In (a) and (c), "inside the NYSE session" now
means from 09:30 ET until an hour after that session's close (17:00 ET on a full
day). It is recorded as a dated append at the end of
research/2026-08-01_b4_skew_activation_pin.md, with a pointer in CAPTURES.md's
options_skew section. The grace was the steward's recommendation (10 of the
first 37 term-stream days disqualified, against 15 with no grace).
The body's two-sided control survives the grace. Re-measured today from
data/options_skew.jsonl's captured_at:
| date | rows | capture time (ET) | under the grace | required |
|---|---|---|---|---|
| 2026-09-14 | 889 | all 18:24 | all past it | must RED (holds) |
| 2026-09-15 | 888 | all 18:03 | all past it | — |
| 2026-09-16 | 884 | 879 at 15:16, 5 at 18:01 | 879 valid, ≥ 400 | must stay GREEN (holds) |
So the gate's window is the close plus 60 minutes, not 16:00 ET, and the controls stand as written.
Unblocked, not approved. This is now an ordinary approval candidate (STEWARD.md §4). The owner ruled on the rule, not on the build.
Triage annotation — 2026-10-19 (annotation only; the body above is untouched) — THE PRECONDITION IS DISCHARGED, AND THE COUNTERSIGN HAS ALREADY CHANGED A SIGNAL DATE
The ruling this row was waiting for landed on 2026-10-15. The owner
countersigned the 2026-09-02 B4 append, clauses (a)–(c), with a 60-minute
grace past the close — in-session plus one hour, 17:00 ET on a full day
(research/2026-08-01_b4_skew_activation_pin.md:262-290, pointer at
CAPTURES.md:158-168). The header's PRECONDITION: not approvable for build until you rule at :8-9 is satisfied. This row is now an ordinary
approval candidate, and the steward's 2026-10-16 carried-forward list already
names it the next one.
Two of this row's own arguments are falsified, both against it.
-
The base rate is 6 of 23 stamped days (26%), not "2 of 14". Measured from
data/options_skew.jsonl, every leg converted to ET and tested against the countersigned 17:00 window:date rows capture instant(s), ET all past the grace? 2026-09-14 889 18:24 yes 2026-09-15 888 18:03 yes 2026-09-16 884 15:16 (879) + 18:01 (5) no — 879 valid 2026-10-05 282 20:43 yes (under the 400 floor) 2026-10-06 882 18:57 yes — unrecorded 2026-10-08 883 19:42 yes — unrecorded 2026-10-16 879 17:59 yes — unrecorded Three whole days above the ≥400-row floor — 10-06, 10-08, 10-16 — are all-post-grace and appear in no register anywhere.
-
"The countersign is free in both directions right now… the window closes the next time a post-close capture lands on a day that clears the row floor" (
:461-465) was already false when it was written. That window closed on 2026-10-06, the next calendar day, nine days before the ruling.
And the clause now bites on live data, within 24 hours of landing. Signal dates recomputed under the pin's own rule (the last date in each ISO week with ≥400 valid rows), with and without clauses (a)+(c):
- ISO W41 → 2026-10-09 either way (10-06 and 10-08 are disqualified, but 10-09 is the week's last date regardless).
- ISO W42 → 2026-10-15 (884 valid) with the clause; 2026-10-16 (879 rows, all at 17:59 ET, 0 valid) without it. That is the first divergence in the series' history — the ruling changed a signal date four days after it was signed, and nothing in the system noticed.
One sentence of the ruling's own rationale is contradicted by the data. The
pin at :286-288 says "Since 2026-10-07 the capture runs on tbx-win at 12:00
PT (15:00 ET, in-session), so the ruling mainly decides how the delay-regime
days count." 10-08 (19:42 ET) and 10-16 (17:59 ET) are both post-cutover
and both post-grace — the tbx-win era has already produced 2 of the 6
all-post-grace days. Worth a dated append to that sentence, not a rewrite.
The defect itself is intact, line for line. grep -c captured_at src/thales/data/captures.py → 0; the field is read only in cli.py (once),
data/options_skew.py (8) and data/resession.py (22). cli.py:4702-4706
(snapshot-skew) still gates on is_trading_day(market_session_date()) and
nothing else, while session_phase — imported at cli.py:12 — is used at
exactly one site, cli.py:2201, the shortability writer. The asymmetry the
row describes is unchanged. capture_qa's options_skew branch
(captures.py:771-800+) remains row-count-vs-trailing-median plus
unparseable-line checks. CAPTURES.md records no post-close day for
options_skew on any of the six dates — clause (b)'s register of record is
still empty.
Two notes for whoever builds it. The free half — writing the six displaced
days into CAPTURES.md per clause (b) — is minutes, and it is six days now,
not two. And this row's own reopen trigger (ii) at :417-418, "any date shows
two distinct captured_at clusters", is fired by ordinary healthy top-ups
(09-04, 09-11, 09-16/17/18, 10-07, 10-12, 10-14) and is unusable as written.
Live state: precondition discharged · blocked on nothing owner-side
· build ~0.25 pd (one captured_at predicate over the ruled 17:00-ET window,
with the 09-14-must-red / 09-16-must-stay-green fixture pair) · the next
approval candidate, and the clause it tests is already changing the record.
ANNOTATION 2026-10-26 (panel run 44) — THE PRECONDITION IS DISCHARGED; AND THE COUNTERSIGNED RULE IS NOW QUOTED BY A SECOND PIN THAT DOES NOT APPLY IT
Still open, not re-proposed. Three facts. The first unblocks the row, the second measures what its absence has cost, and the third — measured against my own first draft of this note — is that the cost is registration wording, not a wrong verdict and not a wrong date in substance. Read fact 3 before acting on fact 2.
1. The blocking precondition was discharged eleven days ago
The header says "PRECONDITION: not approvable for build until you rule on the
2026-09-02 B4 append". That ruling landed 2026-10-15 (CAPTURES.md:172-179):
clauses (a)–(c) are COUNTERSIGNED, with a 60-minute grace past the close — a
signal date's rows are valid only if captured_at falls between 09:30 ET and
17:00 ET on a full session, and "a date whose only capture is later than that
is not a signal date." The row has been approvable for eleven days and is not
approved, and nothing enforces the countersigned rule: grep -n 'captured_at' src/thales/data/captures.py still returns zero hits at HEAD 5c9d335,
re-verified this run. That is this row's whole thesis, now with a countersigned
rule behind it instead of a draft.
2. A second pin quotes the rule as a conjunct and then excludes nothing under it
CAPTURES.md (added by #283, 038cf3b, 2026-10-24) pins the captured-session
unit for the two 60-session clocks — the same block once per stream, at
:377 (options_iv_term) and :464 (shortability), with clause (3) itself at
:383-384 and :470-471 — verbatim:
A file counts toward a 60-session clock only if (1) its date is an NYSE trading day, and (2) its row count is at least 0.5 × the median of the trailing 10 files of the same stream … Where the stream records a capture instant, (3) that instant must fall inside the 2026-10-15 window (09:30 ET to close + 60 min). … Excluded today:
shortability/2026-07-26(a Sunday) andoptions_iv_term/2026-08-07(33 rows vs a floor of 3,801). … Counts through 2026-10-23: shortability 54 … 60th session is 2026-11-02; options_iv_term 50 … 60th is 2026-11-06.
The exclusion list applies clauses (1) and (2) only. Measured from the committed
store, captured_at converted to ET: twelve options_iv_term sessions are
served by an instant outside clause (3)'s window and none is excluded —
08-06 (20:43 ET), 08-26 (17:45), 08-27 (23:32), 08-28 (22:02), 08-31 (19:10),
09-01 (17:41), 09-02 (22:22), 09-14 (18:24), 09-15 (18:03), 10-05 (20:43),
10-06 (18:57), 10-08 (19:42). Two more (09-03, 09-04) carry a valid in-session
instant plus a post-window top-up, and clause (3) — singular "that instant" —
is silent on them, where the pin it generalises from fails a date only if "its
only capture is later than that".
Method check, so this reads as clause (3)'s delta and not as a rival counting method: applying clauses (1)+(2) alone reproduces the published 54 and 50 and the published 11-02 and 11-06 exactly.
Three of the twelve (08-06, 08-27, 08-28) are not oversights: they are the
output of the 2026-09-06 owner re-session ruling (CAPTURES.md:710-720),
which deliberately filed late-evening captures onto the prior trading day by
capture instant. A relabelled row's instant is past that session's close by
construction — that is why it was relabelled. So clause (3), read literally,
disqualifies the product of an earlier owner ruling.
shortability is the mirror image: src/thales/data/borrow_flags.py writes
no capture-instant field at all, so clause (3) can bind on nothing for the
stream whose clock matures first. Yet #282, in the same sitting two minutes
earlier, quarantined shortability/2026-10-16 using "the same job's chains
instant" as the proxy (CAPTURES.md:730). The ruling meant to generalise that
act is silent on the proxy it had just used. Following that precedent instead
gives 54 → 43 and moves 11-02 to 2026-11-17 — eleven post-window
sessions, the same eleven whether the instant is inherited from
data/options_chains or from data/options_iv_term — the inherited instant
is identical from either source, which is the cross-check that makes #282's
unnamed proxy well-defined and turns "the ruling is silent on the proxy it had
just applied" from a hazard into a one-sentence fix. Eleven rather than twelve
because data/shortability has no 2026-08-28 file at all (the store jumps
08-27 → 08-31), so the shortability list is the options_iv_term list minus that
date — consistent from two directions.
A figure of 20 exclusions → 35 sessions → 2026-11-30 was raised during this run. It does not reproduce from either instant source and is recorded here only so it does not survive in the record as a live alternative.
Dates under each reading (NYSE calendar, every session from 10-26 captured):
| reading of clause (3) | shortability | 60th | options_iv_term | 60th |
|---|---|---|---|---|
| A — not applied (what §6 now pins) | 54 | 2026-11-02 | 50 | 2026-11-06 |
| B — applied from the 2026-09-02 field birth | 48 | 2026-11-10 | 44 | 2026-11-16 |
| D — applied, 2026-09-06 re-sessions honoured | 45 | 2026-11-13 | 41 | 2026-11-19 |
| C — applied to every dated file | 43 | 2026-11-17 | 38 | 2026-11-24 |
D is the reading most faithful to the pinned rulings taken together, because applying clause (3) literally to the 09-06 relabels destroys the output of the other pinned owner ruling. Mixed-instant days (09-03, 09-04) are counted as captured sessions on the ground that a date with at least one in-window row is a capture; judging a file by its latest instant would let a post-close top-up destroy a good in-session capture, which is the opposite of what the 10-15 ruling is for.
3. The correction to my own draft — the verdict does not move, and the pinned dates are probably right
My first draft of this note led with "two clock dates are wrong." I measured the thing that would make that matter, and it does not.
The options_iv_term kill criterion (CAPTURES.md:358-362) is "each name's
median-per-day count of distinct captured expiries; if the cross-name
median is < 5 … the stream is RETIRED." The objection to admitting
post-close days is that an after-the-close snapshot loses front expiries and
drags that median down, i.e. toward retiring a moat asset.
Computed exactly as pinned — for each name, the median across its captured days of that day's distinct-expiry count; then the median across names — under the two readings that differ most:
| reading | sessions | cross-name median | names | names below 5 | verdict |
|---|---|---|---|---|---|
| A — clause (3) not applied (what §6 pins) | 50 | 7.0 | 898 | 218 | KEEP |
| C — clause (3) applied to every dated file | 38 | 7.0 | 898 | 219 | KEEP |
The verdict is KEEP under both, by the same wide margin, and the two readings differ by one name out of 898. Per-day, the same statistic is flat at 7.0 on 10-05, 10-06, 10-08, 09-14 and 09-15 — every post-close day — and on each of their in-window neighbours.
Provenance, recorded because the sequence is what a later reader will get
wrong. 7.0 was published by panel run 40 and is quoted inside the
2026-10-24 ruling's own bullet (CAPTURES.md:389-391: "Panel run 40 measured
the stub's effect on the kill statistic: zero (cross-name median 7.0 with or
without it)"). Run 44 re-derived it under both readings as a reproduction of
an already-recorded figure — not a first look at an unmatured criterion, and
no breach of the 60-session pin. The 2026-11-06 computation is therefore a
confirmation of a published number, not a first look.
The front expiry sitting further out is a real property of a post-close capture and a different property from the count of distinct expiries: the front can move out while the count stays seven. So the stream is KEPT under every reading of clause (3), and nothing trades on either clock.
What is actually defective, then, is the wording. Clause (3) was imported
from the skew signal-date pin — where intraday timing is load-bearing for a
rank-IC statistic — into a generic captured-session unit governing two streams,
one of which accepts post-close marks in its own registration
(CAPTURES.md:351-357: "IVs may come from late-session or post-close (staler)
quotes … treat a day's row as an approximate EOD mark … lag consumption by one
day"). Read for intent, clause (3) binds the skew stream and the published
54/50 and 11-02/11-06 are correct. Read for its words, it binds both and the
dates move one to three weeks. The ruling cannot tell you which, and it was
pinned before computation — which is exactly right, and exactly why the
narrowing has to be recorded now rather than discovered in code later.
4. The part that survives at full strength under either reading
research/queue/approved/rad-1-row-adequacy-unit-for-captured-sessions.md
(approved 2026-10-24 for the 2026-10-28 implementer slot) specifies its
negative control as "a weekend-dated file and a sub-floor file are excluded; a
normal file is not." That tests clauses (1) and (2). There is no
instant-window case, so a build implementing two of the three conjuncts
passes its own acceptance test, and nothing anywhere can say which reading it
implemented. A detector that cannot fail is the finding
(ops/RESEARCH_PANEL.md #4) — and this one sits in the spec that both 60-session
clocks now name as their computation.
The panel cannot fix it: research/queue/approved/ is the owner's directory by
path and no agent may write it, and the ruling text is pinned owner-only under
NORTHSTAR Amendment A1 §3. So it is routed, not acted on:
- To the steward, tonight, before Wednesday's build: add an instant-window case to RAD-1's acceptance test, and have the command print the count under each reading with the per-file exclusion reason instead of silently picking one. ~0.05 pd on top of work already scheduled. That keeps the slot on time, keeps the choice with the owner, and makes the disagreement loud.
- To the owner, as the only decision he must make: narrow clause (3) to the streams it was written for, in a dated append that carries the measurements above in front of it. Recorded that way it is a ruling made on the evidence and pinned; discovered after RAD-1 prints a number, it is a choice made with the data on disk. Routing note: his only channel is the Friday digest, and Friday 2026-10-30 falls after Wednesday 2026-10-28 — which is why the recommendation above is "print every reading", not "wait for the ruling".
5. What this changes about this row
Nothing about its diagnosis, which was right in September. It raises the
stakes from "two days passed a perfect QA score" to "the countersigned rule
is now quoted as a conjunct by a second pin that does not apply it, and no
failing check anywhere reads the field." The remedy is unchanged and still
~0.25 pd — a failing check that reads captured_at against the countersigned
window. Recommended for approval at the next triage, on the ground that it
is the only open row whose subject is enforced nowhere and is now load-bearing
for two §6 clocks.
Discharge condition, sharpened against the obvious wriggle-out. The cheap ending — asserting that clause (3) binds only streams with a signal-date rule — is almost certainly the correct one, and asserted bare it would read later as a post-hoc rescue of a published number. So:
C's residue is discharged only when the narrowing append cites, by commit, this annotation — carrying (i) the 38–43-vs-50/54 measurement and the full four-reading date ladder, and (ii) the 7.0-under-both-readings re-derivation with its run-40 provenance. A bare narrowing that does not cite them reopens this as a row.
It dies equally if RAD-1's build filters captured_at against the grace for
both streams and corrects the counts and the §6 dates in the same PR. It
does not die on "the verdict is 7.0 either way": that is measured above and
is why this is a ~0.05 pd wording fix rather than an alarm — not a reason to
leave two readings live on the day a criterion is computed.
Triage annotation — 2026-10-26 (annotation only; the body above is untouched) — ELEVEN DAYS UNBLOCKED, AND NOW LOAD-BEARING FOR THE CLOCK THAT MATURES FIRST
The precondition cleared on 2026-10-15 and this row is still unapproved. Its
header says "not approvable for build until you rule on the 2026-09-02 B4
append"; that ruling landed eleven days ago (CAPTURES.md:172-179, clauses
(a)–(c) countersigned, 09:30–17:00 ET). It is now the cheapest live-biting row
in the tree whose only blocker is a git mv.
Not overtaken, reproduced exactly. grep -c captured_at src/thales/data/captures.py = 0 at HEAD — the field is read only by the
writer (data/options_skew.py), data/resession.py, and once in cli.py.
snapshot-skew still gates on is_trading_day(market_session_date()) alone
(src/thales/cli.py:4793-4795), while session_phase is used at exactly one
site — the shortability writer at cli.py:2243. The asymmetry is intact.
New, and it is why this row now outranks its own filing severity. The
2026-10-24 captured-session ruling quotes the countersigned instant rule as its
clause (3) and pins it twice (CAPTURES.md:384, :471) — but the published
exclusion list applies clauses (1)+(2) only (:389-396), and
src/thales/data/borrow_flags.py writes no instant field at all. So
clause (3) can bind on nothing for shortability, the stream whose clock
matures first (2026-11-02). #282 had to use the same job's chains instant as a
proxy, and says so verbatim at CAPTURES.md:730. Meanwhile RAD-1 — approved for
the 10-28 slot — lists its tests as the two known exclusions plus a weekend file,
a sub-floor file and a normal file (approved/rad-1-…:368-372): no
instant-window case. Wednesday's build would therefore pass its own tests while
implementing two of three conjuncts.
Mitigating, and it should be said plainly: clauses (1)+(2) alone reproduce the published counts (54 and 50) and both maturity dates (11-02, 11-06) exactly, and panel run 44 measured the options_iv_term retire statistic at 7.0 under both readings (218 vs 219 of 898 names below the <5 bar). So this is a wording and enforcement defect, not a mis-dated clock.
Effort ~0.35 pd — ~0.25 for the build, ~0.05 for RAD-1's instant-window
acceptance case, ~0.05 to write the displaced days into CAPTURES.md. Blocked on
nothing owner-side; src/ and tests/ only.