Thales
← research journal
Oct 24, 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 2026-10-24 — the owner's interactive session cleared every open ask: 10 PRs merged, 1 left for Monday, zero questions for the owner

Window: since the 2026-10-23 report. This was an interactive Claude Code session on the owner's Mac, with SSH to tbx-win and broker credentials. It ran on the owner's instructions: "i want this handled by you automatically", "change it then", "change the prompt", and "take care of everything here now". Read: the 10-23 digest (READ), open PRs (READ), the research queue (READ), tbx-win (READ over SSH), broker truth for all three sleeves (READ).

Did

  • merged #276 — #245's main-push-watch retry fix, plus the alert text: a crash now says "the watcher could not look", not "gate bypass". Superseded #245, now closed.
  • merged #277 — the four GitHub backup crons moved an hour later for the 2026-11-02 clock change (trading 16:30/35/40Z, digest 16:50Z), with every doc changed in the same commit. Skew stays at 19:00Z, because 20:00Z is the close under EDT.
  • merged #278 + #279 — STEWARD.md §0b (the NEEDS-OWNER filter, standing rulings, the HANDS lane), ops/HANDS.md, and steward prompt r3, mirrored to the live routine and read back.
  • merged #280 — FRZ-1, superseding #263. Its phrase-scan test was moved to the observability tier, so a memo quoting the phrase cannot fail-close trading. The guard change only tightens.
  • merged #273 + #281 — the beacon reader logs its own failures, and its git calls now pass safe.directory. Root cause of the 10-23 "7 routines with no readable beacon", verified on the box: the repo is owned by BUILTIN\Administrators, the digest runs as thales, and gh is not installed. Monday's box digest is the first proof.
  • merged #268 — the 10-21 report and its LAB-1 + BGS-1 approvals, which the 10-28 implementer slot needs.
  • merged #282 — 2026-10-16 is a skew signal date again. The box's on-time capture (19:00:16Z) was recovered from unpushed/20261016-skew and replaces GitHub's late one. The late rows are quarantined, not deleted. capture-qa passes. Superseded #275, now closed.
  • merged #283 — RAD-1 ruled before computation. A captured session = an NYSE trading day, with rows ≥ 0.5 × the trailing-10 median, and inside the 10-15 instant window. This excludes exactly shortability 07-26 (a Sunday) and iv_term 08-07 (33 rows). Clocks mature 2026-11-02 (shortability) and 2026-11-06 (iv_term), both now in NORTHSTAR §6. RAD-1 is approved, so the 10-28 implementer builds the evaluator.
  • merged #284 — sweep-status --days 60 for momentum + meanrev, from broker truth: 391 append-only status records.
  • Box chores (10-19's ask): boot.py on the box already matched main (SHA-256 d3d295ee…), so no copy was needed. The meanrev reconcile reports "OK — local state matches broker".
  • Closed #274 and re-landed its report as 2026-10-23.md, with its 8 asks marked RESOLVED so the Friday digest does not re-ask them.
  • Standing rulings applied, not asked: no Actions spending (wait for the ~11-01 reset); email stays weekly_digest_only.

Alerts

  • meanrev AVB: asset AVB is not active at Alpaca (a corporate action, most likely). The position is $120 in the negative-control sleeve. Nothing on our side can sell it. The daily sell attempt fails harmlessly until the broker settles the corporate action. No action.
  • VRP status sweep: every lookup failed with a TypeError, because multi-leg orders have no top-level symbol. Fixed in #285 and verified against the live broker: 8/8 VRP orders since 09-04 resolve to FILLED. The same bug would have broken the mleg duplicate-id recovery.

NEEDS-OWNER

(none)

HANDS

(none open)

Carried forward

  • Monday 10-26: merge #270 and #285 (both trading path, both fully tested; Mon–Thu only, never Fri 10-30). After #285, thales run's automatic sweep records the VRP outcomes.
  • Confirm on Monday's box digest that beacon reads work (#281). If any still show "?", its new WARNING lines name the cause.
  • Wednesday 10-28 implementer slot: LAB-1, BGS-1(b), RAD-1 (deadline 11-02), and FRZ-1 is now built.
  • First PST week (from 11-02): confirm no double-runs (AUDIT.md one-off row).