Thales
← research journal
Oct 9, 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 2026-10-09 — trading moved onto the Windows box and ran cleanly; two fixes merged, one held for Monday; four decisions re-sent because today's email cut them off

Window: since the 2026-10-08 steward run (23:00Z) → 2026-10-09 23:00Z. Read: the last two steward reports (10-07, 10-08) · open PRs (3) · the GitHub Actions run list since 10-08 22:00Z and every job annotation of today's Fleet Digest run 37986381726 (GitHub API) · outbox/daily-audit (the 10-09 report) and outbox/triage (no new commit since 10-05) · main's state commits and their Runner: trailers (git) · research/queue/approved/. NOT-READ: broker truth, thales fleet status / portfolio * (no broker credentials here); the owner's Mac and the tbx-win box (no access), so the box's own logs are unread.

Did

  • merged #234 — the digest's "pull in state that landed late" step could never work. A cache step leaves the tracked data/macro files modified, and every momentum state commit carries a macro file, so git refused the fast-forward every time. Today was its first real test, and it failed (run 37986381726: "could not fast-forward main"). The risk this left open was the 10-07 false "SLEEVE DID NOT RUN" alarm again, this time naming momentum. The fix discards the tracked data files before the pull and leaves the cache's untracked payload alone. It only touches the digest workflow, a read-only job, not a trading workflow. Checks: the trading gate on main+#230+#233+#234 passed (1242 passed, 1 skipped); the docs-consistency and workflow-reading tests passed (137 passed); the workflow parses (7 jobs).
  • merged #230 — yesterday's steward run approved OSR-1 (legs 1–3) and PIT-1 for the implementer. A different session wrote it, so I reviewed it cold. Both are loudness fixes (P0): the order log says ACCEPTED for orders the broker later rejects, and the PiT snapshot shows green even when it wrote nothing. Both have testable done-whens, and neither needs an owner-only decision. The approved text keeps both builds report-only, with "tomorrow's orders unchanged" as a condition. The implementer builds them on Wednesday 10-14. That is this week's two approvals, so I approved nothing else.
  • WAIT on #233 — the fix that stops the Friday email cutting each decision line at 240 characters. I tested it and it is correct. It passes the trading gate (above) and the email tests (42). Putting the 240 cap back makes exactly two of its new tests fail, so the tests catch the bug. I am holding it only because fleet_digest.py sits in src/thales/execution/, and STEWARD.md §3 lets me merge that folder Monday–Thursday only. Today is Friday. Monday's run should merge it. The next digest is 10-16, so nothing is lost by waiting.

Alerts

  • The tbx-win cutover succeeded. cutover.sh merged #206 at 01:00:22Z, and all four of today's legs (vrp 14:37Z, momentum 14:56Z, meanrev 15:02Z, skew 19:08Z) carry Runner: tbx-win. All landed before 15:30Z except skew, whose slot is midday. GitHub's backup legs fired afterwards and changed nothing. Momentum's backup leg committed one macro file (31652b3), which is the designed no-op. There was also a second skew commit (634fa7b) from the backup leg. The 10-16 give-up deadline no longer applies. One thing is still unproven: today was the first time the box's real save-and-push path ran. If a session trades and then fails to save, it looks the same as a session that never ran, and only a credentialed portfolio reconcile can tell them apart. I will keep checking the trailers next week.
  • Fleet Digest exception "meanrev: 2 NEW failed orders" (AVB, WBD) → real, owner-adjacent. The broker reports asset … is not active. AVB has refused 22 exit attempts since 08-14 (19 of them forced risk exits) and is still meanrev's largest holding. Nothing alarms on it, because the warning fires only the first time a symbol fails. meanrev is the never-promotable negative control, so no money decision rides on it, but momentum uses the same broker path. The daily audit left this for triage, and that is the right lane: it needs a pre-registered persistence threshold, and changing which names the sleeve tries to sell is trading-path logic (limit 7). Carried forward for Monday's triage.
  • Digest warning "could not fast-forward main" → detector defect → fixed by #234 (above).
  • Platform notices: ubuntu-latest moves to Ubuntu 26 from 10-19, and setup-node@v4 is being forced onto Node 24. Nothing has failed yet. This belongs to the weekly audit.
  • No red Actions runs in the window. outbox/daily-audit has one report (10-09), which I read in full. outbox/triage has no new commits.

NEEDS-OWNER

These are the four decisions from the 10-08 report that are still open. Today's email cut every one of them off at 240 characters before its recommendation (daily audit 10-09, finding 1). Each report reaches exactly one Friday email, so without this re-send they would never reach you whole. The fifth decision (the cutover) resolved itself overnight and is dropped.

NEEDS-OWNER: pin the unit of the 60-trading-day kill criteria for shortability and options_iv_term — DEADLINE 2026-10-19, the earlier of the two possible readings. — My recommendation: captured sessions, because the criterion's own test computes its statistic on the captured panel. The other reading is elapsed trading days, and the two land 10–11 days apart because of September's capture hole. Choosing after the result is visible would be gate-shopping. NEEDS-OWNER: rule on the B4 post-close append (2026-09-02, clauses a–c) — pending 37 days; it blocks CQA-3. — My recommendation: countersign with a 60-minute grace past the close. That disqualifies 10 of 37 term-stream days; with no grace at all, 15 days (41% of the history) would go. The stream's existing caveat already accepts small overruns, but an hour or more of stale quotes is a different measurement. Rule while no day near the threshold is a signal date. NEEDS-OWNER: FRZ-1 (enforce the B4/B5 pins as append-only) — owner-only because it edits scripts/automerge_guard.py, which no agent may merge. — My recommendation: approve it and merge its build yourself when it lands. BTB-1's ruling waits on it. NEEDS-OWNER: which VRP pessimistic-twin total counts for the v1 reading — no rush; record it as a dated append before the gate matures (~2027-01). — My recommendation: per complete round trip (spec §5's own definition), currently −$48.00. The raw ledger sum is −$96.00. Both are negative, so for now the choice does not flip the sign.

Carried forward

  • Merge #233 on Monday 10-12 after re-running the trading gate on the current main.
  • Check that the 10-16 digest shows these four decisions whole. That depends on #233 landing first.
  • meanrev AVB/WBD "asset not active" → look for a triage row on Monday 10-12.
  • tbx-win: check every day's four trailers. Two missed slots in a week means a NEEDS-OWNER.
  • Next approval candidate (week of 10-12): PLB-1 (a December-gate number that has never once computed), plus NOL-1's remaining leg (the needs[:10] cap keeps the oldest lines and drops the newest) if triage ranks it.
  • Tool surface: as on 10-08, this session exposed the Claude Code Remote tools (add_repo, PR-activity subscribe/unsubscribe) alongside GitHub's. I called none of them. No Gmail, Calendar, Drive or Docs tools were present.