Thales
← research journal
Nov 3, 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.

2026-11-03: the momentum sleeve's November selection was voided

Decision: owner, 2026-11-03. The 2026-11-02 selection did not rank the universe, so it does not count as this month's selection. The next run re-selects on complete data.

What went wrong on 11-02

  1. The box did not trade. The pre-trade test gate failed on test_resolve_reference_explicit_path_wins_over_pin, which compared a joined path to a forward-slash string. That assert only runs when results/oos_reference_equity.parquet exists, and the file was committed on 10-31, so 11-02 was the first box run to reach it. Fixed in #316.
  2. The GitHub backup leg ran after the close, with no cache. It started at 23:39Z (18:39 ET). The 13-day Actions block had left no data cache, so it downloaded all 903 symbols from scratch.
  3. Tiingo rate-limited it after 37 symbols. The other 866 went to yfinance. Its end date is exclusive, so those names stopped at 10-30 while the 37 Tiingo names had an 11-02 bar.
  4. Only the 37 names were ranked. The panel's newest bar equalled the decision date, so _augment_to_decision_date added no bar for anyone else, and compute_ranked_signals keeps only names with a row on that date. 37 is below the 50 positions the sleeve holds, so every one of them was selected whatever its momentum.
  5. 63 of the 68 kept holdings were unranked too. compute_diff holds a name with fresh prices and no signal, so the target list grew to 100 names.

No check caught this. The in-process freshness gate looks only at the panel's newest bar, which was current, and the run reported success.

Evidence

Ranks below come from the box's healthy panel (885 names ranked, bars through 10-29):

Names selected on 11-0237
Of those, in the true top 504
Their median true rank436 of 885
Their worst true rank843
True top 10MRNA, SNDK, MU, ILMN, LITE, VLO, MPC, DINO, DELL, APA

The repair

  • data/processed/kelly_ledger.json is restored byte for byte to the commit before the bad run (f75c0b3^). This drops the bad blend base (last_targets), the 3.0 scalar that was never applied to the book (last_construction), and October's holding-period booking, which mixed 10-30 and 11-02 closes. The next selection books October again on one consistent panel.
  • The 11-02 run_summary in data/processed/order_log.jsonl now has selection_completed: false and a selection_voided note. All its other fields are as written, and it is the only line changed.
  • The 66 orders were to be cancelled at the broker before the 11-03 open, which is a separate step outside this commit. Broker truth (thales portfolio orders) shows whether that happened; if they filled, the re-selection trades out of them.

The equity row, the portfolio snapshot and the macro files from that run are real observations and stay as they are.

Why not thales run --force

Rehearsed in a dry run against scratch copies of the state, on the box's panel with the staleness limit relaxed for the scratch copy only (its bars ended 10-29):

Plain --force on the state as it stoodRepaired state, normal run
Selection day without forcenoyes
Turnover cap5% (one day since the bad targets)115% (23 days since 10-01)
Final targets103 names50 names
Names outside the true top 50530
Gross, with October bookednot rehearsed19.6%

--force would also have booked a one-day "holding period" into the Kelly pool. The dry run skips that write, so this comes from reading _update_kelly_ledger_returns, not from the rehearsal.

The 19.6% is a rehearsal figure, and the real run's fresher bars will move it. On either figure the true book is less invested than the 26% it holds today: half-Kelly shrinks after a losing period, and the vol scalar is at its 3.0 cap. This repair corrects which stocks are held, and does not make the book more volatile.

Not changed

  • The meanrev control's orders from its own late run stay queued. It re-selects daily, so its next run corrects them.
  • The gap that let this through was closed separately. See research/queue/built/cov-1-selection-runs-on-a-partial-panel.md.