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.

CUQ-1 — WITHDRAWN 2026-10-30 (same day it was filed). Its premise was false: the question it escalated to the owner had already been decided by the steward on 2026-10-28, and the panel had already filed the true, narrower finding as CSB-1 hours earlier

status: dismissed — WITHDRAWN by its own author, the daily audit, 2026-10-30 · superseded by open/csb-1-clock-evaluator-says-unresolved-after-the-ruling.md · filed in PR #307, withdrawn the same evening


Why this row was wrong

It asserted that options_iv_term's maturity date was contested and that only the owner could settle it, with a hard fuse at 2026-11-06. The date was not contested. research/steward/2026-10-28.md records the steward deciding exactly this question two days before this row was written:

Decided the clause-(3) question #298 escalated, by keeping what is pinned. […] I keep 11-06, the reading registered with the ruling. Choosing the stricter reading now, with the counts already visible, would move a pinned date after its data exists (§0 limit 5). […] Under §0b filter 3 this is a keep-as-is decision, not an ask.

So options_iv_term matures 2026-11-06, NORTHSTAR.md §6 is correct, and there is no owner decision and no fuse. The polarity of this row's central table was inverted: it read the written sources as unaware and the detector as informed, when in fact the written sources carry the decision and the detector is stale.

What was actually true, and where it now lives

thales capture-sessions still prints UNRESOLVED — … belongs to the owner/steward about a resolved question. That is real, and it is a one-line wording fix, not an owner ruling. The research panel filed it correctly as CSB-1 at 14:19Z on 2026-10-30, about five hours before this row — with better scope (propose-only, ~0.1 pd), the right kill criterion, and the right negative control (the printed counts and dates must be byte-identical before and after, or the change has quietly moved a clock). CSB-1 also states what this row failed to: narrowing CAPTURES.md clause (3) is the one move the steward already considered and declined, so re-proposing it would relitigate the gate-holder's two-day-old decision.

Triage should act on CSB-1 and ignore this file. Nothing in it is needed.

The error, recorded because the class matters more than the instance

The standing lesson in ops/DAILY_AUDIT.md is "a red alarm is a question, not an answer." This row took a detector's own words — "belongs to the owner/steward" — as the answer, and never asked whether the steward had already answered. The 2026-10-28 steward report was open in that same session for a different check and the decision sits in plain text in its "Did" list. Checking research/steward/ for NEEDS-OWNER: lines is not the same as reading what the steward decided; a keep-as-is decision leaves no NEEDS-OWNER: line by construction, which is precisely why grepping for one cannot find it.

Cost: a false NEEDS-OWNER escalation reached the owner's phone and was retracted the same evening. The debt that matters is not the row — it is that a wrong escalation on the one channel that interrupts him teaches him to discount the next one.

Carried into the procedure as a check, not as a resolution: before escalating anything a detector attributes to the owner or steward, grep the steward's recent reports for the subject, not for NEEDS-OWNER:. The 2026-10-28 record would have been found by grep -rl "clause" research/steward/.

Also corrected by this withdrawal

This row carried a "correction to my own earlier work" saying the 2026-10-28 daily audit was wrong to call the 11-02/11-06 clocks safe. That retraction is itself retracted. The 10-28 claim was right: both clocks were safe and remain so. The narrow lesson from it survives — verifying that a tool's tests pass is not the same as reading what the tool says — but it produced a false alarm here, not a caught one.


Original text, kept verbatim below for the record

CUQ-1 — a pinned forward clock's maturity date is CONTESTED by its own pinned evaluator, and the only place that says so is the stdout of a command nothing runs. options_iv_term fires its RETIRE check on 2026-11-06 or on ~2026-11-24, and every written source says 11-06

status: open · raised daily audit 2026-10-30 · class: G2 loudness — a pre-registered criterion whose maturity date no reader can learn · judgement: YES · effort ~0.25 pd · horizon: the near date is 2026-11-06, seven days out; the decision must precede it or it is gate-shopping


Plain-language summary for an owner reading one paragraph. One of the two option-data captures, options_iv_term, has a pre-registered kill test: once it has 60 captured sessions, if the typical name carries fewer than 5 option expiries the stream is declared useless and retired. The date that test fires is written in three places as 2026-11-06. But RAD-1 — the evaluator you approved on 10-24 and which was built and merged on 10-28, specifically so this clock would be counted mechanically rather than by hand — disagrees with all three. It counts the sessions two ways, gets 55/60 (fires 11-06) one way and 43/60 (fires 11-24) the other, and then refuses to pick, because choosing moves a pinned criterion's maturity and that is your call. It is right to refuse. The problem is where it says so: only on the screen of anyone who types thales capture-sessions. Nothing mails it, nothing reds, the Friday email does not carry it, and the NORTHSTAR appointment book still reads a flat 2026-11-06. So the clock will mature on a date two readings disagree about by almost three weeks, and the only warning is in a command's output.


The divergence, CONFIRMED-BY-EXECUTION

thales capture-sessions, run 2026-10-30 on main at aa88384:

options_iv_term — ... at 60 captured sessions, cross-name median of per-day
distinct expiries < 5 => VACUOUS, stream RETIRED
  files on disk: 56
  captured sessions: 55/60 (clauses 1+2 — the ruling's own arithmetic)
  60th session lands 2026-11-06 if every session from here is captured
  ...
  ⚠ CLAUSE-(3) DIVERGENCE — clauses (1)+(2) count 55 session(s); the literal
  three-clause reading counts 43 (first-instant) / 41 (every row). The
  ruling's own published counts and dates follow (1)+(2), so the readings
  disagree about when this clock matures. UNRESOLVED — the choice moves a
  pinned criterion's maturity date and belongs to the owner/steward, not to
  this evaluator.
      literal reading (first instant): 43/60, 60th session 2026-11-24
      literal reading (every row):     41/60, 60th session 2026-11-27

Why the two readings differ, which is not a bug in RAD-1. The 2026-10-24 ruling (CAPTURES.md, § options_iv_term) defines a captured session with three clauses: (1) an NYSE trading day, (2) rows ≥ 0.5 × the trailing-10 median, and (3) "where the stream records a capture instant, that instant must fall inside the 2026-10-15 window (09:30 ET to close + 60 min)". options_iv_term does record a capture instant. The capture cron is 19:00Z, and GitHub's shared-runner queue has pushed the actual capture to 17:41, 18:03, 20:43, 22:02 and 23:32 ET on many days — all past 17:00 ET, i.e. outside clause (3)'s window. So clause (3), read literally, excludes about twelve files.

But the same ruling publishes its own counts in the same paragraph: "Counts through 2026-10-23: ... options_iv_term 50, so the 60th is 2026-11-06". Fifty is the clauses-(1)+(2) number. The ruling's text and the ruling's arithmetic cannot both be right, and RAD-1 is the first thing to compute both and notice.

Note the asymmetry that makes this specific to one stream: for shortability clause (3) is vacuous — that stream records no capture instant — so its 59/60 and its 2026-11-02 maturity are not in question. Only options_iv_term is contested.

What every written source currently says

sourcematurity it statesaware of the divergence?
NORTHSTAR.md §6 forward-clock calendar2026-11-06no
CAPTURES.md § options_iv_term (10-15 ruling)"about 2026-11-05"no
CAPTURES.md § options_iv_term (10-24 ruling)2026-11-06no
thales capture-sessions (RAD-1, on main)11-06 or 11-24/11-27, UNRESOLVEDyes — and only here

The buildable part (this row), and the part that is not

Not this row, and not any routine's: which reading binds. That moves a pre-registered criterion's maturity date, so it is the owner's, and it is carried as a NEEDS-OWNER: escalation by the daily audit of 2026-10-30. A routine must not decide it, and deciding it after 11-06 data exists is the gate-shopping the whole apparatus exists to prevent.

This row: an unresolved maturity-date divergence must be loud somewhere a reader actually looks. RAD-1 already computes the warning correctly; nothing carries it. Candidates, cheapest first:

  1. thales capture-qa surfaces any stream whose capture-sessions reading is UNRESOLVED (it already runs daily in skew-snapshot.yml and already reds that run) — the divergence then has a daily home with no new schedule.
  2. The fleet digest's forward-clock line prints the contested pair rather than a single date when the evaluator reports UNRESOLVED. Note the timing trap: the next Friday digest is 2026-11-06 itself, so the digest route cannot deliver this particular instance in time and is the structural fix only.
  3. A capture-sessions --assert-resolved exit code the pre-trade gate does NOT run (a contested research clock must never be able to stop trading — the 2026-10-21 lesson at DAILY_AUDIT.md §0c and PR #264).

Prefer (1). (2) is right but arrives after this clock; (3) only if (1) proves awkward, and never inside the trading gate.

Negative control the build must carry

A stream whose clause (3) is vacuous (shortability) must NOT report UNRESOLVED, and a stream with every capture instant inside the window must not either — otherwise the new alarm fires on the healthy case and trains the reader to ignore it. Both cases exist on disk today, so both are testable without fixtures.

Provenance of this row's own limits

The daily audit of 2026-10-28 wrote that RAD-1 landing on main made "the 11-02/11-06 clocks safe in the strong sense". That was verified the wrong way — the PR's tests were run, the command was not. Running it is what produced this row. 11-02 is safe; 11-06 is not, and was not on 10-28 either.