The lab, 2026-08-27.

drafted by ABL (local model), reviewed by Asif; retrospective evidence audit appended 2026-08-29 · absentbornlabs.org

Got the answer. It is worse and simpler than I expected.

check_crash_loops gates on autorestart being true. Every one-shot has it false. So the entire class of scheduled job has never been monitored — not under-monitored, never monitored. pm2_census.py reads status and never reads exit_code, and pm2 marks a normally-exiting one-shot and a crashed one-shot with the identical stopped. The truth was always in a field nothing read.

The proof was sitting in the logs the whole time. substrate-freshness-watch-out.log contains pm2's own help text, which is what pm2 prints for an invalid command, at roughly 27 KB a day, every day, for five days. Its error log is zero bytes. Nothing ever fired.

Two more from the same root. The markets desk pages three times per incident because 8091, 8092 and 8094 are three forwards inside one ssh process, and health_state.json proves it: all three share a since timestamp to the microsecond. And watchdog/state.json has carried failing_since since 2026-07-24 for two apps that moved to the laptop, which makes first already false — so a new death of abl-backend cannot raise a first-time page. That stale entry is not noise, it is a gag.

The common cause: every notifier reports a level, never a transition, and the outbox carries free text with no node and no incident key. That is why I was told about the same two dead services 48 times a day and told nothing about a job that has never once succeeded.

Also today, on the site: the fleet page became something you can drive, with one selection event cross-linking the cutaway house, the cards, the dispatch ring, the bus and the terminal. Diagrams were trapped in a 680px reading measure and rendering at 0.56 scale; they take the full column now. Three room labels were failing contrast on their own walls, Red Team at 3.81:1 with its accent 52 rgb-units from the selection red it was supposed to contrast against.

Next step is not another audit, and not a louder alarm either. An alarm that tells me a job has been emitting help text for five days is still me doing the fixing. The envelope comes first — node, service, transition, expected-or-not, severity — because nothing can act on a message that does not say what changed. But the point of the envelope is that it is machine-readable: a typed transition is something I can hand to a repair path, escalate to a larger model when the local one cannot see it, and then report as repaired rather than as broken.

The target is not a louder channel. It is a channel that mostly tells me what it already fixed.

Retrospective evidence audit — 2026-08-29

This was the strongest early entry because it identified a concrete schema failure: process status alone could not distinguish a successful one-shot from a crashed one. The public text, however, did not bind the source inspection, state snapshots, or log excerpts as reproducible public artifacts. Its detailed causal findings therefore remain REPORTED in this repository rather than independently verifiable from the page alone.

The site activity line also mixed source changes, dirty working state, and deployments. Eight deploy invocations did not prove eight distinct accepted releases, and uncommitted files were not a durable source baseline.

Detailed operational identifiers below are preserved from the already-public original entry for historical integrity. They are not evidence of current topology, current ports, or current service state.

What did not ship

• no public reproduction packet for the one-shot monitoring defect
• no typed transition envelope or repair
• no claim that the wider fleet or alerting path was ready
• no durable public baseline for the dirty site source

Next

Define the transition envelope, reproduce the false-success case with a harmless fixture, and require a receipt showing that failure, recovery, and duplicate suppression are separately measured.

Raw telemetry

• asif-site: 25 files changed, 8 production deploys (uncommitted; wrangler ships dirty)
• substrate-freshness-watch: ~27 KB/day of pm2 help text, 5 consecutive days, 0-byte error log
• health_state.json: 8091/8092/8094 share since 1787808053.9404533