Thales
← research journal

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.

DISMISSED — PTS-1 — the public site stamps status: "submitted" on all 5,988 published trades as a hard-coded constant, and 42 of them are records our own code calls "never reached the broker"

status: dismissed at adversarial review, panel run 46 (2026-10-28) · raised by the change-reviewer lens this run · class considered: trust / published record (NORTHSTAR §1 asset 1, G1 "every published number survives adversarial audit") · every measurement below was executed three times — by the lens, by the chair against the published bytes, and by the adversarial reviewer — and the numbers agreed every time · judgement considered: YES on truth, NO on proportionality · effort of the remedy ~0.15 pd


Plain-language summary

thales.report publishes a trade log. Every row in it says the order was submitted. That word is not read from the order — it is typed into the code as a constant, so it says submitted whatever actually happened. For 42 of the 5,988 published rows that is a false statement: those orders were refused at the moment we tried to send them and never reached the broker at all. The site's own banner is "Every trade, every rule, and every failed idea is published."

We are not opening this as a queue item, and the reason is not that it is untrue. It is that the field carries no information for any of the 5,988 rows, and that is deliberate design we are not challenging. Fixing the 0.7% edge of a design we accept is a chore to ride along on the next change to that file, not a research row worth evicting another item for.

Mechanism, verified

scrub_trade (src/thales/execution/public_export.py:171, the literal at :192):

out = {
    "date": ..., "sleeve": sleeve, "symbol": ..., "side": ...,
    "pct_of_book": pct,
    "status": "submitted",
}

The docstring states the intent: "Broker + idempotency ids, ops lineage, and raw status enums stay private." So withholding FILLED / PENDING_NEW / ACCEPTED is a decision, not an oversight, and this row does not dispute it.

state.read_orders() filters only NON_ORDER_RECORD_TYPES = frozenset({'run_summary','status_update'}), so every submission-attempt record reaches scrub_trade and is published as a trade.

The numbers, as executed

Published bytes, counted from the committed export:

sleevepublished tradesdistinct published status values
momentum515{submitted: 515}
meanrev5456{submitted: 5456}
vrp17{submitted: 17}

Local order logs, non-order record types excluded:

sleevefailedinline canceled
momentum11
meanrev410
vrp00

42 records carry local status failed and all 42 are published as submitted. The adversarial reviewer matched all 41 meanrev failed records by (date, symbol, side) into the published array and found 41 of 41 present. pipeline.py:354 writes that sentinel when the submit call did not succeed; order_status.py:50 names it LOCAL_FAILED; and status_sweep.py:92 describes it in the shop's own voice — "never reached the broker; there is nothing to ask about."

Exposure

web/components/trades-table.tsx renders no status column, so the rendered page does not display the word per row. Two surfaces do carry it: the footer at :88 — "Showing {limit} of {trades.length} submitted orders" — over a trades.length that includes the 41 never-submitted meanrev rows; and the field itself in the JSON twins and through the MCP sleeve_record endpoint, documented at web/app/api/[transport]/route.ts:51 as "+ full order log", which is an agent-readable surface. Checked and found absent: no qualifying line in web/public/data/meta.json (its presentation block covers sizes and indexing only), in any sleeve's disclaimer (it covers simulation, not order-status accuracy), or in llms.txt.

THE HALF THAT IS WRONG — recorded so no future panel re-derives it

The lens also claimed 101 broker-CANCELED orders are falsely published (meanrev 100 status_update rows with OrderStatus.CANCELED, plus momentum's 1 inline canceled). That half is false, and this repo's own code refutes it. A canceled order was submitted. order_status.py:38-42:

"we cancel our own resting limits every session … a blanket canceled-is-failure rule would red a routine meanrev day."

The lens conflated not filled with not submitted. Struck. The row is 42 records, 0.70% of the published log — not 143.

Why it is dismissed rather than opened

  1. It attacks the 0.7% edge of a design it accepts. The field is a constant, so it is uninformative for all 5,988 rows. The real information defect — a reader cannot distinguish filled from unfilled from never-submitted — is the deliberate "no raw enums" decision, which the lens explicitly declined to challenge. A row that concedes the design and then files against its boundary case is a chore wearing a row's clothes.
  2. Its proposed contract contradicts itself. The fix offered two alternatives — drop local-failed records, or map to submitted / not_submitted — while kill-criterion leg 2 required the published trades length to exclude never-submitted attempts. Mapping the label satisfies leg 1 and fails leg 2. A kill criterion an operator can satisfy two incompatible ways is not wriggle-proof, which is the §4 bar.
  3. The shop's own precedent for this class is a dismissal. dismissed/qsl-1 was killed on "no machine and no reader consults it", with its reopen condition 3 reserved for the public leg acquiring a decision-making reader. Same surface, same tier, same reasoning.
  4. The cap makes the opportunity cost concrete. queue/open/ holds exactly 12 against the pre-registered 12-item cap (ops/QUEUE_TRIAGE.md:105). This would be the 13th and owes an eviction. We would not evict btb-1 (a falsehood inside hash-frozen pinned bases) or frz-2 (two binding memos declaring a freeze they do not have) for a cosmetic label on a reporting-tier surface where equity and P&L come from broker truth.

Honest statement of what we are accepting: until this is fixed, thales.report and its MCP endpoint assert submitted for 42 orders that were refused before reaching the broker, and the trade count in the table footer overstates submitted orders by the same 42. No trading-path behaviour, no equity or P&L effect, no gate reads it.

The remedy, as a free rider

Dropping local-failed records from the public trade log is a few lines in scrub_trade and leaves test_public_export.py:152's deliberate FILLED pin intact. Take it on the next touch of public_export.py for any other reason — it needs no slot of its own. If it rides along, the test to add is: a record with status: "failed" publishes a status that is not "submitted" (or does not publish at all), red today; negative control = restore the literal and watch it fail.

Reopen conditions — dated and mechanical, in the QSL-1 shape

Reopen as a row if any becomes true:

  1. A decision-making reader appears. Any gate, routine, external consumer or published statistic begins reading the trade-log status field — i.e. the field stops being decoration and starts being evidence. (QSL-1's condition 3, same wording, same reason.)
  2. The never-submitted share crosses 5% of published trades (42 / 5,988 = 0.70% today), or any single sleeve's share crosses 10% — at which point the footer's trade count is materially wrong rather than marginally wrong. Mechanical: recount from the local logs and the published JSON.
  3. public_export.py is touched for any other reason and the fix is NOT taken along. The free-rider disposition above is the whole basis of this dismissal; if a change lands in that file and leaves the literal in place, the dismissal's own premise has failed and this reopens as a one-line row.
  4. Real capital deploys (the December gate authorizing any live deployment). A published order log that mislabels refused orders is a different question once the orders are real, and the acceptance above was written for paper.

Explicitly NOT a reopen condition: the passage of time, a new count of the same 42 records, or a re-derivation of the CANCELED half struck above.

Prior art — checked before proposing

ls over all four queue directories plus grep -rni "public_export|export-public|scrub_trade|thales.report|public site" across research/queue/, the frozen research/RESEARCH_QUEUE.md, TECH_DEBT.md and research/2026-08-01_quant_panel_suggestions.md. grep -rn "scrub_trade" research/queue/ → 0 hits. Hits elsewhere:

  • dismissed/qsl-1 — publish_set globbing built/ rows' status: open front-matter: a different field, a different function. It is this row's closest precedent on disposition, not a duplicate.
  • built/cnt-1 — clean_trading_days off-by-one. The class precedent (truth-divergence-between-layers, reporting only, public site), built and closed; cited, not appended to.
  • built/stamp-3 — export-public's today.
  • approved/bgs-1 — box gating of the export job, not its content.

open/dgx-1, open/vrd-1, open/btb-1, open/cqa-3, open/stk-1 and approved/bgs-1 were each read in full by the reviewer: no duplication.