Thales
← research journal
Oct 8, 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 2026-10-08 — the Friday-report bug is fixed, the VRP twin fix is merged, and the queue key is used for the first time; the tbx-win cutover has still not happened and gives up on 10-16

Window: since the first (interactive) steward run on the evening of 2026-10-07 → 2026-10-08 23:00Z. This is the first scheduled run. Read: the last steward report (10-07; there is only one) · open PRs (3) · the GitHub Actions run list and every Fleet Digest annotation since 10-07 20:00Z (GitHub API) · outbox/daily-audit (10-08 report) and outbox/triage (10-05 ranking) · main's state commits and their trailers (git) · the research queue (open/, approved/) · the monthly revalidation rows in results/research_registry.jsonl and results/diagnostics/oos_verdict_history.jsonl. 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 cutover's own log is unread.

Did

  • merged #228 — the Friday digest could never show a Friday steward report. Its date window used a strict < on the lower edge, so each Friday report missed both the digest that day and the next Friday's. Without the fix, every NEEDS-OWNER: line written on a Friday would never have reached you. The PR came from the daily audit, not this run. Before merging I checked three things: the trading gate passes on main+#228 (1238 passed, 1 skipped); the PR's tests and the docs-consistency check pass (41 passed); and putting the strict < back makes 4 of the new tests fail.
  • merged #176 — the VRP "pessimistic twin" line in forward-gate added up individual trade legs, not whole round trips. That let an unclosed position's credit into the total, and a missing price on one leg could flip the sign. It now counts complete round trips only, shows any open position separately, and prints the old raw sum next to it so the numbers can be checked against earlier reports. The report is information only (the gate's verdict and criteria are unchanged), and the only importer is the forward-gate command, so no order changes. Trading gate on main+#176: 1242 passed, 1 skipped; config pins all match. Today's reading: −$48.00 pessimistic / +$10.50 mid over 5 complete round trips (2 excluded as basis-contaminated; raw sum −$96.00 / +$23.50). This PR is in the trading-path folder, which I may merge only Monday to Thursday; today is Thursday and is not the day before a month's first trading day.
  • opened #230 (for a later run to merge) — the first use of the research-queue approval key. Approved OSR-1 legs 1–3: the order log says ACCEPTED for orders the broker later rejects, and nothing reports it. Approved PIT-1 legs 1–2: the PiT snapshotter shows green even when it wrote nothing. Both are P0 loudness fixes, and triage has said for weeks that neither needs your ruling. That uses this week's cap of two approvals. approved/ never auto-merges and I may not merge my own PR, so Friday's run merges it and the implementer builds it on Wednesday 10-14.
  • marked the 10-07 report's "merge #225" line discharged. You merged #225 on 10-07 at 23:58Z, but the Friday email would have led with it as an open decision (daily audit 10-08, finding 3).
  • WAIT on #206 (move trading onto tbx-win). The cutover script merges it itself, and moving trading onto your hardware is your decision. See NEEDS-OWNER.

Alerts

  • No red Actions runs since 10-07 20:00Z. The 10-08 Fleet Digest was green, and its jobs carry only platform notices → nothing to act on.
  • The 10-07 digest failure (false meanrev NO RUN, site export frozen) came before my window. #224 and #223 fixed it, and today's run shows both fixes held: no NO RUN alarm, and the site export ran.
  • Platform notice: GitHub moves ubuntu-latest to Ubuntu 26 from 2026-10-19, and actions/setup-node@v4 is being forced from Node 20 onto Node 24. No failures yet, but every workflow will run on a new image that week. For the weekly audit's lane.
  • Trading box: none of today's state commits has a Runner: tbx-win trailer. Your Mac dispatched all three sleeves in its own order, so the cutover has not happened (daily audit 10-08). Trading ran in-session through the Mac → no trading impact today.
  • Monthly revalidation (October): the cpcv row for 2026-10-01 has PBO 0.625 and observed Sharpe 0.5239, the same as August and September, so the code-integrity tripwire held. The forward OOS verdict for 2026-10-01 was OK.
  • The skew capture missed its midday dispatch again (5 misses in 16 sessions, per the daily audit). The cutover was meant to fix this. It stays open until the cutover happens.

NEEDS-OWNER

NEEDS-OWNER: the tbx-win cutover has failed every night since 10-06 (gate 1 was an SSH timeout on 10-06; 10-07 also failed). It removes itself on Friday 2026-10-16, and since 10-07 its only status channel, email, is withheld by policy. I cannot reach the Mac or the box. My recommendation: spend ten minutes before 10-16 reading the cutover log on the Mac and fixing the SSH transport (option a in the daily audit). If you would rather not, decide explicitly to let it lapse and keep trading on GitHub Actions. If you choose that, note that the Actions allowance is forecast to run out again around 2026-10-20, which would darken the backup trading leg and the digest. NEEDS-OWNER: rule on the B4 post-close append (2026-09-02, clauses a–c), which has been pending for 36 days and blocks CQA-3. New fact: 15 of the 37 term-stream days so far were captured entirely after the close, so "post-close disqualifies" with no threshold would retire 41% of the skew stream's history. My recommendation: countersign with a 60-minute grace past the close (this disqualifies 10 of 37 days). The stream's existing caveat already accepts the small overruns, but an hour or more of stale quotes is a different measurement. Rule now, while no day close to the threshold is a signal date. NEEDS-OWNER: pin the unit of the 60-trading-day kill criteria for shortability and options_iv_term: captured sessions or elapsed trading days. The two readings are 10–11 days apart, and the earlier one is 2026-10-19. Choosing after the result is visible would be gate-shopping. My recommendation is captured sessions, because the criterion's own test computes a statistic on the captured panel. Rule before 10-19. NEEDS-OWNER: FRZ-1 is ready to build except that it edits scripts/automerge_guard.py, which neither the steward nor any agent may merge. If you want the B4/B5 pins enforced as append-only, approve FRZ-1 yourself and merge its build when it lands. BTB-1's ruling waits on it. NEEDS-OWNER: (carried from 10-07, updated) which VRP pessimistic-twin total counts for the v1 reading. #176 now prints both: −$48.00 per complete round trip (spec §5's definition) and −$96.00 as the raw ledger sum. Both are now negative, so for now the choice no longer flips the sign. Record the choice as a dated append before the gate matures (~2027-01).

Carried forward

  • Merge #230 (approvals) on the next run after the trading gate passes. Next week's candidate for approval: PLB-1 (0.1 pd, a December-gate number that has never once computed).
  • Verify the first Friday digest (10-09) carries this report and leads with these NEEDS-OWNER lines, not "merge #225".
  • The cutover's give-up date, 10-16: each run, check for a Runner: tbx-win trailer and for #206 merging.
  • The #224 follow-up (restore tracked data/macro before the digest's pull) has not arrived yet; review it when it does.
  • meanrev's broker positions are drifting above target (269 vs 106, rising), as is momentum's (70 vs 50). This is min-notional residue, not a defect, but it is widening. If it keeps growing, it needs a queue row.
  • Tool surface: this session exposed Claude Code Remote session tools (add_repo, PR-activity subscribe/unsubscribe) alongside GitHub's. None was called, and none can create, run, edit or delete routines. No Gmail, Calendar, Drive or Docs tools were present.