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.

NRB-1 — DISMISSED ON ARRIVAL: "the SLEEVE DID NOT RUN escalation is blind to a never-activated sleeve" — it is not the only machine-readable signal, and the rendered alarm fires

status: dismissed · raised AND killed in panel run 50 (2026-11-16) · class: would-be P0 loudness · never filed as open


What was claimed. fleet_digest.py:116-125 correctly arms the per-sleeve no_run flag on day one via execution.live_start_date: "2026-11-16", but the rollup feeding --fail-on-no-run (:834-838) additionally requires len(equity_curve) > 0. data/state/robo/ does not exist, so if the robo's first run — 2026-11-16 — died, "the only machine-readable outage signal stays silent".

The code reads as described. The live_start_date boundary is correct (today >= start is True on 11-16); :834-838 does carry the equity-curve condition, deliberately and tested at test_fleet_digest.py:224-242.

But the load-bearing claim is false, forty lines above the cited code. fleet_digest.py:810-818:

any_red = fleet_halted or any(not s["ok"] for s in sleeves) or any(
    s["ok"] and (s["data"].get("no_run") or ... ) for s in sleeves)

No equity-curve condition. So a missed robo run on day one still:

  • sets fleet.red = True;
  • yields digest_exceptions(f) → "fleet red (halt, kill-switch, sleeve error or no-run)" (:1728-1729);
  • renders 🔴 DID NOT RUN today (trading day, no run record) in the text digest (:1298-1299) and the HTML twin (:1432-1435) — the exclusion test itself asserts this (assert mr["data"]["no_run"] is True # still rendered as not-run);
  • and on a Friday that exception string ships in the owner's one email.

What is actually silent is far narrower: the named no_run_sleeves list and the --fail-on-no-run exit code (cli.py:3898). "The only machine-readable outage signal" was the adjective carrying the severity, and it does not hold.

On the heartbeat question, which cuts both ways and settles neither. GitHub Actions is dark (heartbeat-monitor.yml last success 2026-11-06; every run 11-10…11-14 failure — the third monthly billing refusal), so the robo's self-expiring grace ([ "$(date -u +%F)" \< 2026-11-18 ]) would red nothing from 11-18 because nothing runs. But even with Actions healthy a red heartbeat reaches no person: _alert.yml:84-100 annotates, writes a job summary and raise SystemExit(0) under weekly_digest_only. So the heartbeat neither covers this nor fails to cover it — it is not a channel to a human at all. Both "kill because the heartbeat covers it" and "keep because the heartbeat is dark" are wrong, and both appeared in the candidate.

Residue, stated precisely: for a sleeve that has never succeeded, --fail-on-no-run returns 0 indefinitely and the exception line says "or no-run" without naming the sleeve. One-line fix (and (len(curve) > 0 or due)). Worth an ops/HANDS.md line if someone is already in the file; not worth a registry row and a deflated-Sharpe hurdle.

Reopen conditions (any one)

  1. A sleeve's missed first run is shown to have reached no human and no record on the day it happened. That is the direct falsification of this dismissal and needs no argument — one digest text plus one beacon/outbox check.
  2. digest_exceptions or any_red is ever narrowed to require an equity curve — the rendered alarm goes with it and the gap becomes real.
  3. A second sleeve is activated while --fail-on-no-run's exit code is used as a gate by anything that acts (a workflow that halts, pages or blocks on it), making the unnamed-sleeve gap consequential rather than cosmetic.