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.

SWC-1 — the steward's "Carried forward" calendar reaches no channel: the Friday digest gathers only the headline and NEEDS-OWNER: lines, so the two 60-session clock maturities and the monthly decay check are in no email

status: DISMISSED 2026-10-30, panel run 47 (killed by this run's own adversarial review, not by capacity) · raised by the queue-and-clocks lens, panel run 47 · class: G2 loudness · effort would have been ~0.3 pd


What was measured, and it is all true. fleet_digest.py:648 (_gather_steward_week) collects, per steward record in its window, the # headline and every line beginning NEEDS-OWNER: — nothing else. EXECUTED: grep -c '^NEEDS-OWNER:' over research/steward/2026-10-2*.md returns 0 for every record from 10-23 through 10-28 (the last prefixed line is 10-22). The clock maturities (2026-11-02, 2026-11-06) and the ~11-01 monthly decay verdict check sit under ## Carried forward in research/steward/2026-10-28.md and in #304, unprefixed. So today's digest — the owner's only email — names none of them, and the next digest is 2026-11-06 itself.

Why it is dismissed anyway

The harm does not bite. Nothing on 11-02 or 11-06 requires an owner decision: the shortability clock is a floor whose consumption gate gives onto a hash-frozen battery pre-registration that does not exist, and the options_iv_term check is a mechanical, report-only computation whose evaluator shipped on 10-28. NORTHSTAR.md §6 pins both clocks WITH named evaluators — which is G5's deadline ledger working, not failing. The steward is the designated reader under the 2026-10-07 delegation and runs every weekday, and ops/RESEARCH_PANEL.md §3#5 makes the clocks this panel's daily watch, which is the channel that actually carries them.

What is left once that is subtracted is "the owner would not learn of a maturity if the steward went dark for several weekdays" — and that residue is already owned by three open rows: rlv-1 (the digest's dark-routine count forgets a miss), dgx-1 (the digest's cadence gate), and dlv-1 (an unmerged steward record drops from every digest). Filing a fourth variant of the same theme would be the padding NORTHSTAR.md §2 warns about.

Reopen conditions (either one)

  1. A NORTHSTAR §6 clock reaches maturity with no evaluator run recorded anywhere within 3 trading days of its date. That is the condition under which the missing channel would have cost something real. The next two tests of it are 2026-11-02 and 2026-11-06 — so this tripwire is live within a week of being written, which is the best kind.
  2. A steward outage of ≥3 consecutive weekdays spans a §6 maturity. Precedent exists (the six-routine outage 2026-10-12→10-14), so this is not hypothetical; it is simply not what happened.

Not reopened by: another Friday digest going out without the calendar in it. That is the dismissed state, not a change to it.


ANNOTATION 2026-11-02 (panel run 48) — two of this dismissal's load-bearing premises were falsified by #310 the day after it was written; the conclusion survives on its other legs, and reopen condition 1 is LIVE THIS WEEK

Still dismissed. This is a record, written because ops/RESEARCH_PANEL.md "What matters here" #2 says future panels evaluate dismissals rather than skip them — and a future panel re-evaluating this one would otherwise reason from two sentences that are no longer true.

The falsified premises

This file's "Why it is dismissed anyway" rests partly on (:26-30, verbatim):

"The steward is the designated reader under the 2026-10-07 delegation and runs every weekday, and ops/RESEARCH_PANEL.md §3#5 makes the clocks this panel's daily watch, which is the channel that actually carries them."

Both clauses were false within a day of being written. PR #310 (1491885, owner's 5× cadence cut, merged 2026-10-31 — this file is dated 2026-10-30) put the steward on Thursdays only (0 15 * * 4 PT, confirmed from the live control plane) and this panel on Mondays only (0 6 * * 1 PT). The "channel that actually carries them" went from five observers a week to one, and ops/RESEARCH_PANEL.md still self-describes as daily at seven sites, which is now filed as research/queue/open/pcs-1-pipeline-diagrams-contradict-the-live-crons.md.

Why the dismissal nonetheless stands

Its other legs were re-checked today and hold:

  • The harm still does not bite. Verified by running the pinned evaluator from this container: shortability 59/60 (60th lands 2026-11-02), options_iv_term 55/60 (60th lands 2026-11-06), pinned statistic: not due — not computed. Shortability remains a floor onto a hash-frozen battery pre-registration that does not exist; options_iv_term remains a mechanical, report-only computation. Neither needs an owner decision.
  • The residue is still owned by open/rlv-1, open/dgx-1 and open/dlv-1, so a fourth variant would still be the padding NORTHSTAR.md §2 warns about.
  • Both reopen conditions are intact and are not being weakened.

Reopen condition 1 is live this week — and it was partly discharged today

Verbatim (:40-44): "A NORTHSTAR §6 clock reaches maturity with no evaluator run recorded anywhere within 3 trading days of its date. The next two tests of it are 2026-11-02 and 2026-11-06."

  • 2026-11-02 (shortability): discharged by this annotation. Panel run 48 ran thales capture-sessions on the maturity date itself — 0 trading days — and the output is recorded above and in the run's beacon/panel marker. One operational correction that makes this repeatable: two of this run's own lenses reported the command NOT-RUN for want of polars, and runs 47 and 45 recorded it as runnable only "on the box/Mac". It is runnable in a routine container: pip installs under /usr/local/lib/python3.13/dist-packages, which /usr/bin/python3 reaches, so /usr/bin/python3 -c "from thales.cli import app; app(['capture-sessions'])" works (as does /usr/local/bin/thales). python3 -m thales.cli <cmd> exits 0 printing nothing — there is no __main__ guard — which is how a lens concluded the command was dead.
  • 2026-11-06 (options_iv_term): NOT yet dischargeable, and the only routine inside its window is the panel. The 60th session lands Friday 19:00Z, after that day's digest (08:30 PT) and after the steward's Thursday 11-05 slot, which will read 59/60 and print "not due". Routine slots: triage 11-03, implementer 11-05 17:02Z, daily audit 11-05 22:07Z, steward 11-05 23:00Z, then panel 2026-11-09 14:00Z — one trading day after maturity, inside the 3-trading-day margin. So the 2026-11-09 panel run must run thales capture-sessions and record the verdict, or this condition fires on its own first test. Carried forward in run 48's report and beacon rather than scheduled as a follow-up run, which the owner forbade on 2026-10-30.

Also recorded, from the same adversarial pass

The panel's own §5 cadence criterion was examined this run and no row was filed: its remedy ladder reads "drop to weekly or on-demand", so the second rung is unspent and the criterion remains live; ops/QUEUE_TRIAGE.md counter (c) already names its executor. A panel filing a governance row about its own cadence is the NORTHSTAR.md §2 degenerate shape, and it was killed on that ground.


ANNOTATION 2026-11-09 (panel run 49) — reopen condition 1 was tested twice this week and did NOT fire; both clocks were evaluated inside the margin, and the second one now has a verdict

Still dismissed. This is the discharge record for the tripwire this file wrote one week before it could be tested: "A NORTHSTAR §6 clock reaches maturity with no evaluator run recorded anywhere within 3 trading days of its date … The next two tests of it are 2026-11-02 and 2026-11-06." Both tests are now complete and the condition held — but only because a routine was told to hold it, which is the honest caveat recorded at the end.

Test 1 — shortability, matured 2026-11-02. Did not fire.

Evaluator run on the maturity date (panel run 48) and again on 2026-11-05 by the steward, which wrote the maturity into prose with a decision — research/steward/2026-11-05.md:15: "This report is the written record of that maturity … Decision, no owner action: keep capturing, and write no HTB battery now. The 60 sessions are a floor, not a deadline." Re-measured today: 64/60 captured sessions, 65 files, clock MATURE, one exclusion (2026-07-26, clause1_not_a_trading_day). Recorded within 0 trading days.

Test 2 — options_iv_term, matured Friday 2026-11-06. Did not fire, and the criterion is now ANSWERED.

Run here on 2026-11-09, one trading day after maturity, inside the 3-day margin. thales capture-sessions, verbatim:

options_iv_term — CAPTURES.md § options_iv_term — at 60 captured sessions,
cross-name median of per-day distinct expiries < 5 => VACUOUS, stream RETIRED
  files on disk: 61
  captured sessions: 60/60 (clauses 1+2 — the ruling's own arithmetic;
                            last counted 2026-11-06)
  clock MATURE — the criterion is due
  pinned statistic over 60 captured sessions, 898 names:
      cross-name median 7.0 distinct expiries/day (floor 5); 219 name(s) below it
  KEEP — above the floor

Verdict: KEEP. The free option feed is deep enough to carry a term structure; the stream is not vacuous and is not retired, so the nightly write stays and B4's input is untouched (the two streams share a cron but B4's pinned primary reads options_skew's skew field, so a RETIRE would not have touched it either). The clause-(3) divergence printed alongside (literal readings 48/60 and 46/60, 60th session 11-24 / 11-27) is the one RESOLVED 2026-10-28 in favour of clauses (1)+(2), as NORTHSTAR §6 pins; it is printed for comparison and changes nothing. Verified a second way, bypassing the CLI entirely. The statistic was recomputed in this run straight from the 61 committed data/options_iv_term/*.parquet files — group by (date, underlying), count distinct expiries, take each name's median across its captured days, then the cross-name median — giving 898 names, cross-name median 7.0, 219 below the floor: identical to the pinned evaluator's output to the digit, from independent code. It also matches the 7.0 measured by panel run 40 on a 34-file panel and by run 44 under both clause readings (218 vs 219 of 898). Four reads agree, and this one is at the pinned maturity rather than as a projection. The verdict is not close to its threshold: 7.0 against a floor of 5, so no plausible dispute about the exact 60-session set changes it.

Both §6 rows are therefore discharged on their merits and neither needs an owner. Carried forward, not filed as a row: §6's table still lists both clocks as pending and CAPTURES.md § options_iv_term carries the criterion without its answer, so the KEEP verdict currently lives in this annotation, in this run's beacon marker and in the session report — not in the two documents a future reader consults. §6 is a living section and CAPTURES.md is the authoritative ledger; writing either is the steward's move or the owner's, not this panel's (ops/RESEARCH_PANEL.md rails: propose only). The steward's own 2026-11-12 plan already carries "Read the options_iv_term kill statistic if the 11-09 panel has not" — it has been read, and what remains is to record the answer in those two places.

The caveat this file should carry about its own tripwire

Panel run 48's marker instructed this run, in capitals, to run the evaluator today precisely so condition 1 would not fire. The condition therefore measured a routine obeying an instruction left for it — not an independent channel. Run 43 already recorded that this class of condition "is partly under the panel's own control and so is a weak tripwire", and that reading is reinforced, not weakened, by today's pass: the five-observers-a-week channel that condition 1 implicitly assumed became one observer a week after the 2026-10-30 cadence cut, and the margin was one trading day wide. A pass under these conditions is evidence that the hand-off worked, not that the channel is robust. The next maturity with this little margin should be treated as an expiry, not a routine item.

Three mechanical facts behind why this clock kept needing a reader

Verified independently in this run, and recorded here because each explains a recurrence rather than a one-off:

  1. Today was the only reader that could have satisfied condition 1. The condition allows 3 trading days. Computed with the repo's own calendar: 2026-11-06 + 1 td = 2026-11-09, +2 = 11-10, +3 = 11-11, and the steward's next scheduled run, 11-12, is +4. So the steward's own plan ("the 11-09 panel or my 11-12 run reads it") was never two options — only this slot was inside the margin. Condition 1 did not fire by one trading day.
  2. The KEEP branch has no record site, by construction. CAPTURES.md attaches the record duty to the RETIRE branch only: "the capture is recorded VACUOUS … and the stream is RETIRED — annotation appended here, the write removed from the run." A KEEP is pre-registered to be computed and then to name nowhere to be written. Same shape as dismissed/scr-1's "and that gets recorded", which named no site either.
  3. The authoritative capture ledger is the one record file no routine can land. CAPTURES.md appears in neither ALLOWED_PATTERNS nor DENY_PATTERNS in scripts/automerge_guard.py, so it is an unrecognised path and the guard fails closed. Executed: automerge_guard.py --files CAPTURES.md --statuses CAPTURES.md:modified → DENY: CAPTURES.md is not a record file (records-only fast path), while AUDIT.md, RESEARCH.md, TECH_DEBT.md and the three writable queue directories all auto-merge. So writing a verdict into the ledger needs a human-merged PR; writing it into AUDIT.md or RESEARCH.md does not. That is the mechanical reason the ledger lags its own criteria, and it is not proposed as a change here — the guard failing closed on an unrecognised path is the guard working.