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.

SIT-1 — the public site's next export republishes a RETIRED sleeve as a live daily-trading negative control, the sentences that assert it are hardcoded in the generator, and the test that should catch it grades the exporter against its own constant

status: open · raised panel run 50 (2026-11-16) · class: published-record truth (NORTHSTAR asset #1 — trust) + a detector that cannot fail · judgement: YES · effort ~0.2 pd, observability tier, steward-mergeable · horizon: decidable immediately; it is authored on the next successful export


Plain-language summary. The public site thales.report publishes a card for each trading strategy. The meanrev strategy was retired on 2026-11-13 and its broker account deleted. The code that builds the site does not know what "retired" means — its two sibling programs do, it alone does not — so the next time the site is built it will publish meanrev again, with no end date, under the heading "Negative control: a strategy our own tests REJECTED, run live on purpose." The site's own summary file, rebuilt from scratch every day, also states in plain words that three strategies "paper-trade daily" and that "one sleeve is a deliberate negative control." Both sentences stopped being true on 2026-11-13. The test that is supposed to prove the site lists every strategy compares the site against the site's own list, so it cannot ever fail.

Scope — read this before anything else

The robo sleeve's absence from the site is NOT part of this row. That is a recorded, dated deferral: research/2026-11-13_robo_sleeve_spec.md §11 defers the robo's site page to "before the interim read". Including it would be padding. This row is three things, one mechanism:

  1. meanrev republished as live,
  2. three hardcoded sentences that assert a composition that no longer exists,
  3. a completeness test that is true by construction.

What is true (verified at HEAD f606336; every claim with the command)

1. The exporter has no retired check; its two siblings do. src/thales/execution/public_export.py:411-413:

for name, sl in SLEEVES.items():
    if name not in PUBLIC_META or not Path(sl.config_path).exists():
        continue

Compare src/thales/execution/fleet_digest.py:50-52 and src/thales/cli.py:4005, which both read or sl.retired. Three consumers of one registry; two were updated at retirement, one was not. git show --stat f606336 confirms public_export.py is not in the activation commit.

2. The asserting sentences are GENERATED, not static. public_export.py:527 writes into llms.txt: "Three strategy sleeves paper-trade daily on separate broker paper accounts … One sleeve is a deliberate negative control." :494 writes into meta.json: "Mean-reversion negative control: full record." Both are rebuilt with a fresh generated_at on every daily export.

3. The site is STALE now, and becomes AUTHORED-FALSE on the next export. This timing matters and the row would be wrong without it. git log -1 f606336 → committed 2026-11-14 00:04 PT. The last commit touching web/public/data/fleet.json is 2026-11-13 17:45 PT (#333). So the bytes currently served predate the retirement — they are stale, not yet a false statement authored after the fact. The next successful export (the box's 08:30 PT fleet-digest task) regenerates fleet.json, meta.json and llms.txt with a new timestamp and republishes meanrev as live. That is why this is cheap now and expensive later. Current published bytes, read directly:

$ python3 -c "json.load(open('web/public/data/fleet.json'))"
generated_at 2026-11-14T01:45:17Z
momentum  flagship          live_since 2026-03-20
meanrev   negative_control  live_since 2026-07-13   <- no end date, no retired field
vrp       experiment        live_since 2026-07-13
(no robo — deferred, not this row)

4. The detector cannot fail — demonstrated by negative control, not argued. tests/test_execution/test_public_export.py:98: assert set(summary["sleeves"]) == set(PUBLIC_META) — where PUBLIC_META is the same constant the exporter filters on at :411. The change-reviewer lens registered a fifth live sleeve ghost in SLEEVES, re-ran the production export, and the assertion PASSED with exported = [momentum, meanrev, vrp] against registered live = [momentum, vrp, robo, ghost]. The next assertion (:103-105) is a hardcoded file-set literal, so it catches a REMOVAL and never an ADDITION. pytest tests/test_execution/test_public_export.py -q → 15 passed on current main.

5. Nothing forces the site to be considered when a sleeve is added or retired. tests/test_utils/test_sleeves.py:280-330 (test_registry_coverage_forcing_functions) checks the workflow file, .gitignore, SLEEVES.md, CLAUDE.md and the strategy class — not PUBLIC_META — and is @pytest.mark.observability, i.e. deselected from the pre-trade gate. grep -in "public\|site\|web/\|report" ADDING_A_SLEEVE.md returns one hit, about a broker-account guard. The "Retiring a sleeve" section added by this very commit (ADDING_A_SLEEVE.md:126-146) has four steps — registry phase, workflow/session/heartbeat/env removal, positions, config note — and no step touches the published record.

Prior art — checked, and the nearest neighbour argues FOR this row

grep -ril "public_export\|PUBLIC_META\|export-public\|thales.report" research/queue/ → dismissed/pts-1 (a hardcoded trade status column), dismissed/pbc-1 (69 names vs a 50-name label), dismissed/qsl-1, plus unrelated hits in open/cov-2, built/bgs-1, built/cnt-1, built/stamp-3. All are value-level clarity on sleeves that are present; none is about a sleeve being published that no longer trades.

The nearest neighbour is the frozen archive, research/RESEARCH_QUEUE.md:2359 (2026-08-20): "public site renders the halted meanrev sleeve as healthy", killed to "a fold-in on the next public_export.py touch". It is not a bar, on three grounds:

  1. Different defect — that is a missing field (no halt-adjacent key); this is a missing filter plus three hardcoded sentences.
  2. New evidence, three days old — the retirement is 2026-11-13; nothing before it could have raised this.
  3. The fold-in disposition has empirically failed. public_export.py was touched twice in this delta — 63d6fd7 (#332) and 724638d (#333), both 2026-11-13 — and the fold-in did not happen either time. That is evidence against choosing the same disposition a third time, and it is the single best reason this is a row rather than another deferred fold-in.

Done-when

  1. public_export.py's sleeve loop gains a retired check matching fleet_digest.py:50-52 — or publishes meanrev with an explicit end date and a retired role label. Do not simply drop it: the site's promise is that every rule and every failed idea is published, so erasing a retired sleeve's record would be the opposite defect. Preserve the record, correct the tense.
  2. The three hardcoded strings (:494, :527 ×2) become derived from the registry — sleeve count and roles computed, not written.
  3. test_export_public_full_surface's first assertion is re-pointed from PUBLIC_META to SLEEVES filtered by retired, with its negative control shown in the PR: register a live sleeve absent from PUBLIC_META, see the test go RED. A test that passes against the bug is decoration.
  4. The durable half — ADDING_A_SLEEVE.md's add- AND retire-checklists each gain a published-record step, and test_sleeves.py:280-330's forcing function gains PUBLIC_META. Without this the same defect recurs on the next sleeve, and that is what makes this a row rather than a patch.

Pre-registered kill criterion

This row dies if either:

  • (a) public_export.py gains a sl.retired check matching fleet_digest.py:50-52, AND the three hardcoded strings become derived from the registry, AND test_export_public_full_surface is replaced by an assertion that can fail, demonstrated by the negative control in done-when 3. (Three conjuncts deliberately: fixing only the filter leaves llms.txt asserting three daily-trading sleeves.)
  • (b) the owner rules that the published record is allowed to lag the registry — in which case that ruling is written into ADDING_A_SLEEVE.md's retirement checklist in the same edit, so the next reader is not left to rediscover it.

Effort / horizon

~0.2 pd. Observability tier, no trading-path code, steward-mergeable. Decidable immediately; the defect is authored on the next successful export.