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.

AMM-1 — DISMISSED — the derived Actions-minutes meter that AUDIT.md's register says "needs its own queue/open/ item"

status: dismissed 2026-11-02 at panel run 48's adversarial pass (raised by the queue-and-clocks lens the same run) · class: G2 loudness · effort would have been ~0.1 pd to file, ~0.5 pd to build


What was proposed. AUDIT.md:126 closes its Actions-allowance row with: "A working meter therefore has to be derived — sum started_at→ completed_at per job from /actions/runs/{id}/jobs, round each job UP to the whole minute, group by billing month — or be given account scope … That is code, so it needs its own queue/open/ item; this row can only stop the third exhaustion being a surprise." No such item exists. Two allowance exhaustions in six weeks (2026-09-19 → 09-30, and 2026-10-19 00:15:59Z → 10-31), both discovered after the fact from refused runs.

It is buildable — that is not why it dies. The adversarial reviewer probed /repos/MaksimDan/thales/actions/runs/{id}/jobs from a routine container with $GH_TOKEN and got per-job started_at / completed_at, so the derived meter is well under a person-day. The 2026-11-01 weekly audit had already proved the two obvious alternatives dead: the account meter is HTTP 403 from a routine ("sessions are bound to their configured repositories") and /actions/runs/{id}/timing returns billable.UBUNTU.total_ms: 0 on every run sampled.

Why it is dismissed

  1. At threshold there is no action anyone here may take. "Never pay for GitHub Actions" is a standing owner ruling (ops/STEWARD.md §0b), and no policy anywhere pre-commits what gets switched off at 80% of the allowance. By this shop's own doctrine — ops/QUEUE_TRIAGE.md:114, "counters with no committed action at threshold are decoration" — a meter without a pre-committed response is decoration. The queue is not the place to put a number nobody may act on.
  2. The harm has largely been engineered away. Since 2026-10-07/19 the always-on Windows box carries trading, the skew capture, the fleet digest and the public-site export. An allowance exhaustion now costs the backup legs and the beacon relay, not a trading day or a capture day — so it is no longer in the irreplaceable-data class (NORTHSTAR.md §1 asset #2).
  3. An audit row instructing that a queue item exist is not evidence of harm. Filing it would move a known gap from one list to another, which is the records-maintenance drift ops/RESEARCH_PANEL.md warns about.

The better-shaped alternative, recorded so it is filed instead of the meter

The refusal is already machine-readable: the 2026-11-01 weekly audit read "The job was not started because recent account payments have failed or your spending limit needs to be increased" off run annotation 41128716879. A detector on that annotation string closes the entire "discovered after the fact" gap at a fraction of a meter's cost, and it fires on the first refused run rather than on a forecast. If anyone files in this area later, file that, not the meter.

Reopen conditions (any one)

  1. A third exhaustion costs a trading or capture outcome — not merely a backup leg, a beacon relay or an auto-merge run. That moves it from loudness to the moat and it reopens as a P0-adjacent row.
  2. A pre-registered "at X% of the allowance we disable Y" policy exists, so the number finally has a committed action at threshold.
  3. The box stops carrying the time-sensitive jobs (trading, skew, digest, public export) and GitHub Actions becomes the primary runner again, which restores the harm reason 2 removes.
  4. The owner asks for allowance forecasting, which makes it a request rather than an inference.

Not reopened by: another exhaustion that costs only backup legs, or by AUDIT.md:126 continuing to say a queue item is needed. That sentence is the dismissed state, not a change to it.