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.

BCN-1 — the liveness beacon is gated on job SUCCESS in five workflows, so a post-trade reporting failure withholds the "did it run" signal

status: dismissed by PANEL run 24 (2026-09-16), at its own adversarial review, before reaching the queue · class: loudness (G2) / detector coupling · never occupied a row in queue/open/ — filed here so the measurements are not re-derived and the reopen valve is visible to ls


Plain-language summary for an owner reading one paragraph. Every scheduled job pings healthchecks.io at the end of its run; that ping is the only thing in the system that proves a job ran at all. In five workflows the ping is gated on the whole job having succeeded. In the two sleeve workflows the reconcile step runs after trading is done and state is committed — so a failure in a post-trade report withholds the liveness signal, and the sleeve reads as "never ran" when in fact it traded normally. That is what happened to meanrev from 09-11: a false drift alarm (the defect PR #171 addresses) held thales-meanrev DOWN for four days. A detector defect disabled a detector. This is dismissed anyway, because the owner was in fact paged on all three red days through a channel that does not depend on the beacon, and what is left is a design question only the owner can answer.

Mechanism — verified at HEAD 7fd0d74

if: ${{ success() }} guards the healthchecks ping at:

workflowline
paper-trading-meanrev.yml192
paper-trading.yml374
paper-trading-vrp.yml137
fleet-digest.yml119
heartbeat-monitor.yml83

In the sleeve workflows thales portfolio reconcile runs if: always() after Persist sleeve state, so the ordering is: trade → persist → commit → report → beacon. A red report reaches back past three completed steps and suppresses the beacon.

Why it is dismissed — the backstop is real, and it fired

paper-trading-meanrev.yml:203-210 runs the shared _alert.yml on always() && needs.trade.result != 'success' && != 'skipped'. It emails independently of both the beacon and the fleet digest, and it fired on 2026-09-11, 09-14 and 09-15. Separately, fleet_digest.digest_exceptions (fleet_digest.py:1171-1172) lists no_run_sleeves as an exception that forces a same-day email even under the new weekly cadence. So the owner was never actually un-pushed — which also corrects the claim in PR #171's memo that "a genuinely missed meanrev run would have looked identical." It would not: the two states are identical only on the healthchecks channel, not on the fleet's.

What remains is a genuine design ruling with no defect behind it: should a liveness beacon mean the process ran or the process ran clean? Decoupling them trades the false-red this fixes for a new false-green risk (a job that runs, fails, and still reports alive). That is an owner call, not a bug, and minting a row for it at a 12/12 queue would displace an incumbent that has one.

Reopen condition

Reopen as a row if any of these becomes true:

  1. a beacon is withheld on a day when the _alert path also fails to deliver (the backstop is what carries this dismissal — if it lapses, the dismissal lapses with it);
  2. any sleeve's beacon stays down ≥3 consecutive sessions for a cause outside the reconcile window-edge class that PR #171 addresses;
  3. the owner rules on the semantics and wants the decoupling implemented — at which point this is a build item, not a question.

Adjacent ground, do not duplicate: DGX-1 (open) holds the digest-gate blindness, including the missed=None "no readable beacon" leg; FOS-2 (dismissed) holds the per-sleeve streak split.


ANNOTATION 2026-09-25 (panel run 28) — reopen condition (1) AND (2) FIRED; the dismissal has lapsed by its own terms, and the row is STILL not the right instrument

Read this paragraph if you read nothing else. This file was dismissed on 2026-09-16 because the thing it worried about — the liveness ping being withheld — was backstopped by a separate alert email that "fired on all three red days". Between 2026-09-21 and 2026-09-25 that backstop stopped working, so the condition this file pre-registered for its own resurrection came true. But the fix this file proposes would not have helped, because the way it came true is one the file did not anticipate. Both halves of that are recorded here so the next reader does not have to choose between honouring a pre-registration and doing the sensible thing.

What fired, and the evidence

GitHub has refused every Actions job in this repository since the 09-19/09-20 weekend over an account billing failure (research/2026-09-21_actions_refused_ every_job_over_billing.md, days 1 and 2–4). Against this file's "Reopen condition" list:

  • (1) "a beacon is withheld on a day when the _alert path ALSO fails to deliver" — TRUE, verified by execution. On the 2026-09-24 momentum run (run 36042412357) the GitHub jobs API returns exactly two jobs: trade (107777545034, failure, 18:36:28→18:36:33) and alert / email (107777572703, failure, 18:36:33→18:36:36). The backstop job was created and refused in three seconds, same as everything else. Not inferred from the incident memo — read off the API this run.
  • (2) "any sleeve's beacon stays down ≥3 consecutive sessions for a cause outside the reconcile window-edge class that PR #171 addresses" — TRUE. All three sleeves, four consecutive sessions (09-21, 09-22, 09-23, 09-24), cause = account-level billing refusal, which is plainly outside that class.

Condition (2) stands on its own and needs no reading of (1).

Objection considered and closed. One reading is that "the _alert path" means "the owner hears about it somehow", and GitHub's own failed-run notification is outside Actions. Two things defeat it. First, this file's own text defines the path concretely — "paper-trading-meanrev.yml:203-210 runs the shared _alert.yml … It emails independently of both the beacon and the fleet digest" — and that workflow is the refused job above. Second, the day-1 incident inventory records that the operator's inbox on 09-21 contained exactly three Thales messages, all healthchecks DOWN notices, on a day when roughly ten runs failed. Whatever GitHub's own run-failure mail is doing, it is empirically not arriving.

Why this does NOT become a row

The reopen conditions were written to detect "the backstop lapsed, so the beacon coupling now matters". The backstop did lapse — but because the entire Actions band went down, not because the backstop is unsound. Decoupling the ping from success() in those five workflows would have changed nothing during this outage: the beacon step never executed at all, because the job that contains it was never started. The condition's letter is satisfied by a mode its fix does not address.

This is recorded, not quietly resolved, because reinterpreting a pre-registered condition after it fires is the gate-shopping this shop treats as a cardinal sin. The honest statement is: the tripwire fired, the dismissal's stated justification no longer holds, and the remedy it names is still not worth a slot. If the owner wants the semantics ruled on (should a liveness beacon mean the process ran or the process ran clean?), that ruling is reopen condition (3) and converts this into a build item. Nothing here forces it.

What this instance is actually evidence FOR

Not the beacon coupling — the reading gap. This condition became true on 2026-09-21 and was first evaluated today, four days later, by a panel lens brief that had dismissed/ added to it by hand. In between, the daily audit wrote two long memos (09-21 and 09-24) about the alarm layer inverting without once consulting the file that had pre-registered this exact tripwire five days earlier. That is dismissed/rdr-1-no-reader-for-dismissed-reopen-conditions.md reopen leg (b), and it is annotated there.