Thales
← research journal
Nov 5, 2026raw markdown ↗

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.

Steward 2026-11-05: the 11-02 guard and the FOS-3 alarm fix are merged; the box traded every day this week except Monday, whose cause is fixed; one queue approval; zero questions for the owner

Window: from my 2026-10-30 run (beacon 23:15Z) to 2026-11-05 23:30Z. This is the first weekly run (cadence cut, #310) and covers five trading days, 11-02 to 11-05, plus the 10-30 evening. Read: my last two reports (10-29, 10-30, both on main) (READ) · outbox/daily-audit tonight's report, both addenda and the 10-30/10-31 closers (READ) · outbox/triage 11-03 report in full (READ) · all 7 beacon/* branches (git, READ) · main's commits since 10-30 with their Runner: / Box-Result: trailers (git, READ) · the Actions run list from 11-02 23:42Z onward (GitHub MCP, READ; earlier in the window I used the daily audit's reading) · all 3 open PRs: diffs and their merged result (GitHub MCP + git, READ) · ops/HANDS.md, research/queue/open/ and approved/ (READ) · the monthly rows in oos_verdict_history.jsonl and research_registry.jsonl (READ) · thales capture-sessions and thales verify-config --all-sleeves, run here on main (READ). NOT-READ: broker truth, thales fleet status / portfolio * (no broker credentials) · the Mac and tbx-win (no access) · Fleet Digest run annotations (only run conclusions were read) · the routine control plane · the live thales.report.

Did

  • merged #318 (COV-1, built by the owner's 11-03 session). A selection now refuses when fewer than 90% of expected names have a bar on the price panel's newest date, or when fewer names are ranked than the sleeve holds. A refusal exits non-zero and the next session retries. This closes the 11-02 failure mode, where momentum picked 37 names out of 903. It mattered most for meanrev, which selects every day and was unguarded until this merge. Trading-path class, merged on a Thursday that is not the day before a month's first trading day. What tomorrow's orders do differently: nothing on a normal panel. The box's panel scored 98.1% coverage on 10-29 against the 0.90 pin, and the new run-summary fields are report-only. Triage recommended that I merge it (#319). It adds a keystone pin and moves no existing one.
  • merged #320 (FOS-3, implementer). The exit-4 "new de-risking failure" alarm can no longer be hidden by the previous day's failed buy. It changes the run's exit code only, never orders. The daily audit measured a remainder, DRO-1: a VRP open is a sell, so a failed open can still mask a failed close. That hole predates #320, because the old side-blind set masked it too. The audit recommends merging anyway, and I agree. DRO-1 arrives with #321.
  • Checks before merging (main + #318 + #320 + #321 merged together): trading gate (the box's command, -m "not observability and not research" --ignore=tests/test_lab) 1335 passed, 1 skipped; observability tier 424 passed; docs-consistency + the touched tests 65 passed; verify-config --all-sleeves clean. No engine, portfolio, strategy or feature file changed, so the byte-identical backtest is untouched. Publish scrub: 0 redactions in the 5 changed research/ notes. Squash-merged, each merge pinned to the head SHA I reviewed. Neither PR had a commit from this run.
  • #321 (daily audit 11-05) made mergeable, not merged. It annotates the COV-1 row in open/, and #318 moved that row to built/. I merged main into #321 and moved the annotation along with the row. The resulting tree is identical to the one I tested. The merge commit is mine, so the 11-12 run merges it. It carries DRO-1 and two HANDS items, the outbox/triage relay restore and the CLAUDE.md queue-delay sentence. Neither has a deadline.
  • Approved REF-1 (#323): the decay judge must refuse instead of quietly falling back to the backtest curve that every backtest overwrites. It is a P0 fail-closed fix ahead of the December gate. It lands in approved/ when the 11-12 run merges it, and the implementer builds it 12-05. That is after the row's 12-01 horizon, which is acceptable: the gate cannot be decided before 12-10, and the pin has been committed since 10-31. This is the only approval this week.
  • Authored CSB-1 (#322): thales capture-sessions no longer prints "UNRESOLVED" about the clause-(3) question that I decided on 10-28. Only the wording changes. Counts and dates are identical (59 / 47 / 45 and 11-06 / 11-24 / 11-27), and a negative control reds. The 11-12 run reviews it from scratch.
  • The shortability clock has matured (63/60 captured sessions on 11-05, first mature 11-02, per thales capture-sessions). This report is the written record of that maturity, which triage noted existed only in a beacon commit message. Decision, no owner action: keep capturing, and write no HTB battery now. The 60 sessions are a floor, not a deadline, so continuing to capture loses nothing. A battery written now would be written with the borrow-flag counts already visible. Those counts also show almost no hard-to-borrow names in the R1000 universe: 2, not the "0 of 889" that BTB-1 notes still appears at five places. So a screen on that flag would barely bind. If the owner ever wants the test, the pre-registration is his frozen text to write. Correcting BTB-1's false premise (a dated append) carries forward.
  • options_iv_term reaches its 60th session tomorrow (59/60 tonight), on the 10-28 reading, which stands. The kill statistic (median distinct expiries, retire if below 5) can first be computed after Friday's 19:00Z capture. The 11-09 panel or my 11-12 run reads it.
  • Decided myself under §0b filter 3, so none of these is an ask:
    • RWG-1 / meanrev's 278-order cleanup on 11-03 against the 300-order cap (triage raised it as NEEDS-OWNER). Keep as is. #318 now refuses the 11-02 shape that produced the cleanup. meanrev is a paper negative control that can never be promoted. If the cap ever blocks a cleanup batch, the system halts trading and retries the next day; nothing irreversible is lost. The batch is back to 134 orders today (order_log.jsonl).
    • meanrev's 11-02 record (255 targets on the partial panel, all orders cancelled unfilled, no positions taken): leave as recorded. Rewriting a live sleeve's ledger is owner-only, and with zero fills there is nothing to correct.
    • COV-1 leg 4 (may a backup leg that starts after the close select at all?). The owner's 11-03 session decided it: unchanged. The daily audit's new measurement says the backup legs now start after the close most days (median delay about 4h40m, none under 2h). That reinforces the case for #318's refusal but does not change the ruling's reasoning: with complete data, orders queue for the next open, matching the backtest's next-open timing. Not reopened.
    • FIN-3 (the decay monitor prints OK while live Sharpe is −1.33). Not approved. The substantive fix edits the retire rule of a pinned gate in the middle of the gate, which §0 limit 5 forbids. The additive version could not land before the 12-01 monthly verdict through the implementer, so it would change the vocabulary while the December result is in view. The go-live gate's own Sharpe-lower-bound criterion already fails honestly, and the interpretation memo frames the read. Revisit after December.
    • C6 (go-live rung ladder) and C8 (buy-or-skip), both with 12-01 fuses per triage. No action before December. No real money can move without a gate PASS plus the owner's own go-live decision, and the gate cannot pass in December: 102/126 days, and the live-Sharpe lower bound is −4.86 against a pass bar of > 0. So the ladder can still be registered before any real-money numbers exist. C8: skip, which is no spending.
    • PCS-1's schedule half (the implementer on the 5th often runs before a Thursday steward can approve what triage ranked on the 3rd). Keep the crons. I work around the ordering by approving the week before, and the cost is latency, not loss.
  • Standing rulings applied: AVB's failed exit continues (owner 10-24: "No action"). Its streak was broken by the 11-02 non-session, so on 11-03 AVB counted as "new" and produced the box's failed (trade) result. That is explained, not a new event. The Actions billing block is over (first green 11-01 05:47Z), so the "never pay" ruling needed no use this week.

Alerts

  • Box: 4 of 5 trading days fully on the box. 11-03, 11-04 and 11-05: all three sleeves, the site refresh and skew carry Runner: tbx-win. On 11-02 the box's site and skew tasks ran but trading did not. Its pre-trade gate failed on a path-comparison test that first ran that day (fixed in #316 by the owner's session). GitHub's backup legs traded about 7h late, after the close, which is the incident #318 now guards. One miss with its cause fixed, so not a HANDS item.
  • Red runs since 11-02 23:42Z (Actions list): Fleet Digest 11-02 (the no-run detector, firing correctly) and Skew Snapshot 11-03 01:29Z (capture QA, firing correctly on the PiT-snapshot hole). Both belong to the 11-02 incident and are explained. Earlier in the window I relied on the daily audit's count: three reds, all on 11-02.
  • Monthly re-validation (Mac): a new cpcv row dated 11-01 with PBO 0.625 and observed Sharpe 0.5239, identical to 09-01, so the code-integrity tripwire held. The OOS verdict row for 11-01 is present (OK, n=98, live Sharpe −1.33). The ~11-03 NEEDS-OWNER condition from my 10-30 report did not fire: the Mac is alive.
  • Daily audit INCIDENT (beacon/triage's relay had been deleted since 10-26): the audit fixed and verified the beacon half itself. The outbox half is a HANDS item in #321.
  • Beacons: every routine due this week moved. The research audit's next run is 2027-01-01.

NEEDS-OWNER

(none)

HANDS

(no new items from me. #321 carries two from the daily audit: restore outbox/triage's relay workflow, and correct CLAUDE.md's queue-delay sentence.)

Carried forward

  • 11-12: merge #321 (records + HANDS), #322 (CSB-1) and #323 (REF-1 approval) after re-testing. Read the options_iv_term kill statistic if the 11-09 panel has not. Consider approving VRD-1 together with DRO-1 so they are built as a pair before the 12-05 implementer slot: the daily audit measured that fixing VRD-1 alone arms DRO-1's masking hole.
  • BTB-1: a dated append correcting the "0 of 889 hard-to-borrow" premise (true count 2).
  • 12-01: the first trading day of December (momentum alpha selection, the first selection under #318) and the oos-monitor's monthly verdict. 12-10: the earliest date the December gate can be decided.
  • Tool surface: this session exposed the Claude Code Remote tools (add_repo, PR-activity subscribe/unsubscribe), a PDF viewer, and two failed finance connectors (BigQuery, Slack) alongside GitHub's. I called none of them. No Gmail, Calendar, Drive or Docs tools were present.