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.

DMC-1 — two operating documents still promise an alarm the owner switched off a month ago, and one of them is the checklist he reads before going away

status: open · raised panel run 50 (2026-11-16) · class: records truth / NORTHSTAR C0 "never spend trust" · judgement: YES, narrowly — and this row is one fifth of what it was proposed as · effort ~0.1 pd, docs + one test, steward-mergeable · horizon: ops/ABSENCE.md's own header dates it before real capital deploys in December

Read the narrowing first. This was proposed as a loudness row about detection latency for a dead trading box. The adversarial reviewer killed that claim and it stays killed — the reasons are recorded below so no future panel re-derives it. What survives is smaller and different in kind: two documents give the OWNER false instructions about an alarm channel.


Plain-language summary. Until 2026-10-07 an outside service (healthchecks.io) would notice if a scheduled job stopped reporting and send the owner a message. He turned that notification off — a deliberate decision, recorded. Two documents never caught up. One is the checklist he reads before going away, and it tells him to make sure those very notifications are on, calling them "the only alarm that reaches a phone on a beach." The other contradicts itself eighteen lines apart: it states the service sends nothing, then states the same service is the only alarm that survives a total outage. The fix is to correct both and add one test so they cannot drift back.

What is true (verified at HEAD f606336, verbatim)

  • ops/ABSENCE.md:47, in the pre-departure checklist: "Confirm healthchecks.io push/mobile notifications are ON — beacons are the only alarm that reaches a phone on a beach."
  • RUNBOOK.md:83-90, the authoritative note: "2026-10-07 (owner, interactive session): healthchecks.io sends NOTHING. The project's only notification channel (email) was removed and the account's periodic email reports were turned off … The 10 thales-* checks remain and are still pinged … so they are a passive dashboard."
  • RUNBOOK.md:92-97, eighteen lines below that note, unamended: "Because it alarms on ABSENCE, from healthchecks' own infrastructure, it is the only alarm that survives the case where no job of ours gets a runner at all."
  • ops/ROUTINES.md:235-237 agrees with the note, not with the claim: "a missed ping alarms nowhere."
  • ops/ABSENCE.md's own header: the page "must be revised before real capital deploys (December)" — about two weeks out.

dismissed/dsp-1 (panel run 25, annotated run 35) already indicted the adjacent line ops/ABSENCE.md:38 as part of the 2026-10-02 beacon-deletion cluster. That cluster is discharged — RUNBOOK.md:102-113 now strikes all four deleted rows with dates and reasons. Line :47 was never touched, and it concerns the later and wider 2026-10-07 channel removal, not the 10-02 deletions. DSP-1 is not reopened by this row.

What was PROPOSED and KILLED — recorded so it is not re-derived

The original row claimed a composed three-layer failure leaving a dead trading box unreported for up to seven days. The reviewer killed it and verified why:

  1. The "~7 days" figure counted only duty-bound readers. The routines that read the repository are the panel (Mon 14:00Z) and the daily audit + steward (Thu 22:00Z / 23:00Z). The longest gap between any routine reading the repo is Thursday night → Monday morning, ≈3.5 days, not seven. The seven-day number was a true statement about duties and a misleading one about detection.
  2. The proposed remedy — "give the Monday panel a pinned box-liveness duty" — is a seventh member of the family this shop has killed six times (swc-1, trp-1, clk-1, rdr-1, cuq-1, bcn-1): a duty in prose, pointed at a routine, with nothing that fires if the routine skips it.
  3. The precedent cited for that duty is itself rotted text. ops/RESEARCH_PANEL.md "What matters here" #5 reads "The clocks are yours daily … you are the only routine that reads the queue every day" — false since the 2026-10-30 cadence cut moved the panel to Mondays (ops/ROUTINES.md:61, 0 14 * * 1). Invoking stale text to justify a new duty is the rot argument running backwards.
  4. dismissed/bcn-1's 2026-09-25 annotation already ruled on this exact mode and recorded that "the remedy it names is still not worth a slot"; dismissed/dgx-1 and dismissed/rlv-1 hold the digest-side half.

None of that is reopened here. What is left is not a latency row.

Done-when

  1. ops/ABSENCE.md:47 corrected — it must not instruct the owner to rely on a channel that was removed. Replace with what is actually true: the checks are a passive dashboard; alarm-on-absence lives in the fleet digest's DARK/no-run detection and the steward.
  2. RUNBOOK.md:92-97 replaced by one composed paragraph that answers, in one place, "if tbx-win stops, what notices, how soon, and who is told?" — written as the correction to the contradiction, not as new machinery. It should name a reader that is neither on the box nor inside GitHub Actions.
  3. One tests/test_docs_consistency.py assertion, with its negative control stated: no document under ops/ or RUNBOOK.md may assert that healthchecks.io notifies, alarms, pushes or reaches anyone while RUNBOOK's 2026-10-07 note stands — and re-adding the Email integration (deleting that note) must make the assertion RED, not silently pass. This is the leg with an actor: a detector, not a duty, and it is what stops the drift recurring.

Pre-registered kill criterion

Dies when grep -ri "healthcheck" ops/ RUNBOOK.md returns no line asserting that the service notifies, alarms, pushes or reaches anyone, AND test_docs_consistency.py carries an assertion that fails if such a line is reintroduced. Mechanically checkable; no judgement required.

It also dies if the owner re-enables the healthchecks Email integration, which makes both sentences true again — in which case RUNBOOK.md:83-90's note is struck in the same edit.

Explicitly NOT in scope

Nothing here proposes re-adding an alarm, un-pausing a check, un-deleting a beacon, or flipping notifications.email. Those were four deliberate owner decisions with recorded reasons and they stand. This row corrects two documents that describe the world as it was before them.

Routing note

ops/*.md and RUNBOOK.md are not on ops/STEWARD.md §3's trading-path list, so the steward may author and merge this directly rather than waiting for the 2026-12-05 implementer slot. Given ops/ABSENCE.md's December deadline, that is the cheaper route. It may equally be discharged as an ops/HANDS.md item — but note that "fold in on the next touch" demonstrably failed twice for public_export.py inside this same week's delta, so if it is routed informally it should carry a dated default.