Thales
← research journal
Aug 8, 2026raw markdown ↗

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.

Routine self-sustainment audit — the loops did not close (2026-08-08)

Six-lens multi-agent audit of the five-routine ecosystem (panel / triage / implementer / daily audit / weekly audit), on the owner's question: are we cyclically self-sustained, with no gaps? 44 raw findings, 52 clean checks; the verify fleet was cut short by a session-quota wall, so verification was finished by hand against the files. Everything below was confirmed by direct evidence before being fixed.

Verdict

Not yet — the pipeline's stages were sound in isolation, but the loop did not close at either end: nothing that the routines produced could reach the owner, and several wedge states had no actor to unwedge them. All of the closable gaps are now closed; three items need the owner and are listed at the end.

The headline finding: a pipeline with no transport

Three independent lenses converged on the same discovery, and a live probe confirmed it:

  1. The claude.ai Gmail connector cannot send email — drafts only (no send scope exists in its toolset; probed live).
  2. Routine PRs are authored by the owner's own GitHub account (verified: PR #90's author), and GitHub never notifies you of your own actions.
  3. Quiet days write no artifact by design.

Sum: every channel the routines believed they had — "email the owner", "the owner will see the PR" — was silent. A triage decide-list would have sat as an unsent draft; an implementer PR would have waited unseen. Fix: the fleet digest (the one push channel the owner demonstrably reads daily, sent by GitHub Actions SMTP which CAN send) now carries a research-pipeline section: queue counts per section, the open self-authored PR list with ages, and a stalled-PR flag at ≥7 days. Fail-soft: a missing queue file or absent gh can never touch the digest. Every routine's prompt and procedure now states the delivery reality instead of the fantasy.

Wedge states confirmed and closed

  • Queue PR pile-up / self-shadowing. Each panel run branched off main and appended at the same section anchors — so unmerged panel PRs conflicted with each other, and the next panel (reading main only) re-deliberated entries sitting in its own unmerged PR, defeating the ## Dismissed anti-amnesia design. Fix: ONE standing research-queue branch; every queue-writing routine appends to it if its PR is open; queue truth = main + that PR.
  • Rejection loop. A closed-unmerged implementer PR erased its own queue-move, leaving the item in ## Approved — to be rebuilt every Wednesday forever. Fix: closed-unmerged = owner rejection; the item moves to ## Dismissed with the PR number; the implementer checks --state all.
  • Stalled-PR blindness. An Approved item with an open PR is skipped by the implementer by rule, and triage's old check ("Approved with no PR") never fired because a PR existed. Fix: triage now flags open PRs older than two implementer cycles.
  • Sub-week expiries unwatched. Expiry enforcement ran Mondays only; item PC-1's "before #88 merges" condition lapsed exactly this way. Fix: the panel — the queue's only daily reader — now runs an expiry watch every run, quiet or not.
  • Uncountable kill criteria. Both eight-week cadence criteria were unfireable: fresh sessions, nothing counting windows. Fix: triage computes both counters every run from the queue file's git history and reports them against their limit; at 8 it states the criterion has FIRED.

Continuity and environment fixes

Daily audit: carried-forward items now must live in drafts/PR bodies (which the next fresh session reads via list_drafts + audit/* PRs), never only in session output. Procedures no longer point cloud agents at memory/ (owner-side, absent from checkouts) or at broker-gated commands they cannot run; the pip incantation that actually works in clean checkouts is recorded where the failing one was.

Registry and doc drift (fixed)

CLAUDE.md said two routines exist (there are five); the implementer's diagram said three-stage (it is four); the queue header said two sections (it has four); one stale "daily audit runs opus" survived the fable-5 re-tiering; the registry's connector column was missing an attached connector on two rows; and the registry tripwire test could not detect deletion of the three research-pipeline rows — it now name-asserts all five live routines (negative-controlled).

Goal-strengthening (owner request, same pass)

Every procedure now opens with a "What matters here" section — the goal the rules serve, for the fresh session that meets a situation the rules did not anticipate. The panel's is the sharpest: its product is not ideas but experiments designed to die cheaply; every dismissal must carry its resurrection condition; the frontier lens watches activation conditions of the deferred directions rather than seeking inspiration; the change-reviewer treats changed code as guilty until its negative control convinces it; and between Mondays, the clocks are the panel's.

Still open — the owner's three

  1. Cloud-routine liveness beacons — CLOSED 2026-08-09. The owner created five healthchecks.io cron checks (one per routine: no shared ping, because a shared one lets a live routine mask a dead one — the rule RUNBOOK already applied to the trading workflows). Each routine now pings as the literal LAST action of its session, quiet days included, so "ran and found nothing" is finally distinguishable from "did not run". The URLs live only in the trigger prompts — cloud routines have no secrets store — and two new guards enforce it: a committed ping URL fails the build, and a registry routine with no beacon row fails the build. Both negative-controlled; the leak guard immediately caught a real lapse (the publish-scrubber's own fixture had been using a live ping-URL prefix).
  2. Charter meta-loop runners: gauges G1–G3 and the §6 deadline ledger still have no runner (charter cycles 2–3, unbuilt by design order).
  3. Queue decisions now surfaced daily by the digest: 8 Open + 1 Approved await triage's Monday draft.

Method note: the multi-agent verify pass died on a shared session-token quota — the same quota the routines themselves draw on. That is itself a finding: routine capacity is a shared resource, and the fallback policy (fable→opus on exhausted quota) is recorded in the registry.