2026-11-19: a fresh site can be a false one — four days of an authored-false public record, with every instrument green
Found by: the daily audit, 2026-11-19. Class: published-record truth (NORTHSTAR asset #1) + a monitoring gap in the audit's own check 10. Still open at the time of writing — the fix is authored but unmerged.
What is published right now
web/public/data/, regenerated by the box's 08:30 PT export and committed as
6ac92bc today, carries generated_at: 2026-11-19T16:30:24Z and these three
sleeves:
| sleeve | role published | live_since | end date |
|---|---|---|---|
| momentum | flagship | 2026-03-20 | — |
| meanrev | negative_control | 2026-07-13 | none |
| vrp | experiment | 2026-07-13 | — |
meanrev was retired on 2026-11-13 and its broker account deleted
(f606336, #336). meanrev.json still carries the present-tense tagline "A
strategy our own tests REJECTED, run live on purpose." llms.txt — the file
written for other models to read — still asserts "Three strategy sleeves
paper-trade daily on separate broker paper accounts" and "One sleeve is a
deliberate negative control." Neither sentence has been true since 11-13.
The robo sleeve, live since 11-16, is absent; that part is a recorded,
dated deferral (research/2026-11-13_robo_sleeve_spec.md §11) and is not
part of this note.
The only trace of the retirement anywhere in the published bytes is that meanrev's curve stops at 2026-11-13. Nothing says why, and nothing says it stopped.
Mechanism — one registry, three consumers, two updated
src/thales/execution/public_export.py:411-413 iterates SLEEVES and filters
on PUBLIC_META membership and the config file existing. It has no
sl.retired check. Its two siblings do: fleet_digest.py:50-52 drops retired
sleeves at gather, and cli.py does the same for thales fleet status —
verified from this container, which prints meanrev | retired | retired and
3 live sleeve(s) / cap 3. The activation commit did not touch
public_export.py.
The panel found this on 2026-11-16 and filed it as SIT-1 (PR #337), with the
timing stated precisely: at that moment the served bytes merely predated
the retirement. Its prediction has since fired. The exports of 11-16,
11-17, 11-18 and 11-19 each re-authored the claim with a fresh timestamp, so
the record moved from stale to authored false, which is the more expensive
state and the one the row said to avoid.
Why nothing caught it — the part worth keeping
Every instrument pointed at the public site read green through all four days.
- The export job ran and committed daily. Green.
- The audit's check 10, half 1 ("was the data authored?") asks whether
generated_atkeeps advancing. It advanced every day. PASS — because the defect was being re-authored. - The audit's check 10, half 2 ("did the host serve it?") is answered by
UptimeRobot, the one monitor outside GitHub and outside tbx-win. Its
2026-11-16 weekly mail reports 100.000% uptime, 0 incidents for both
thales.reportandthales.report/data/meta.json. PASS. test_export_public_full_surfaceassertsset(summary["sleeves"]) == set(PUBLIC_META)— the exporter graded against the same constant it filters on. SIT-1 demonstrated by negative control that registering a fifth live sleeve leaves it passing.
So the three checks nominally covering the published record are freshness, reachability and a tautology. A site that is fresh, reachable and wrong fails none of them. Freshness was being read as a proxy for truth, and under a re-authoring defect freshness is actively the wrong signal: the greener it reads, the worse the state is.
That is the generalisable lesson, and it is the twin of the standing one. "A red alarm is a question, not an answer" has always had a mirror — a green you cannot believe is worse than a red — and this is a green that was not merely unbelievable but inverted.
What was done
ops/DAILY_AUDIT.mdcheck 10 gains a third half: the published sleeve set must equal the registry's non-retired set, retired sleeves must carry an end date, andllms.txt's prose counts must match — with the three commands. Freshness and reachability are explicitly marked as not covering it.- The standing code fix is not written here. It belongs to
SIT-1(PR #337, the row) and PR #338 (the owner's, 2026-11-17), which adds astatusof live/retired to every sleeve, renders the retired card last and puts meanrev's page in the past tense. #338 has been open and unmerged since 11-17 only because GitHub Actions is in its fourth monthly billing refusal, so auto-merge is dark and merging is by hand.
The residual, stated plainly
Until #338 merges, every 08:30 PT export re-authors the claim. The cost is not
hypothetical: llms.txt exists so that other models read the record, and the
site's own promise is that "every trade, every rule, and every failed idea is
published." Publishing a retired strategy as live is a failure of exactly the
asset the project says it is building. The correct repair is the one SIT-1
names — preserve the record, correct the tense — never deletion.