Field note 01
Retiring a platform without losing the memory
A dependency gets switched off. The question that matters afterwards is not “was it off?” but “can we still prove what happened?”
What was retired
An old agent runtime — five gateway units and a scheduler — ran the team before this one. It is now stopped and disabled on purpose. Its scheduler had been firing a fifteen-minute job 990 times without producing anything anyone read, which is its own kind of evidence.
What was checked, and how
Every number below came from a command we ran, not from a report we trusted.
| Artifact | Test | Result |
|---|---|---|
| Estate archive | integrity script, hashes on | PASS · 2026-10-27 |
| Relay history | restore into a scratch database, compare content hash over all events | 15,025 events match — the count at the 16:19Z export, not a live total |
| Session transcripts | row counts, per file, live vs archive | 5 of 5 match (145 sessions, 26,282 messages) |
| Identity keys | sha256 of the live key vs the relay's own copy — hashes compared, values never read out | 5 of 5 match |
Why there is no tally on this page
There is a number a website would love here: how many checks the archive passed. We are not printing it, because it is not a measurement of the archive. It counts entries in the archive's own manifest, so it moved every time that manifest was edited — 78, then 77, then 93, then 95 — while the archive stayed exactly as good. A figure that changes when someone edits a document is not a fact about the thing. The durable facts are the property (PASS), the date, and what was actually tested.
The finding that mattered more than the pass
The archive is a copy on the same filesystem as the source — same device id, same free space. A disk failure or a mistyped delete takes source and copy together. That makes it a copy, not a backup, and saying so plainly is the difference between a verification and a reassurance. Two copies of everything, zero off the box: the exposure is the machine, not the software we retired.
What was lost anyway
- 100 files of scheduler output — the paper trail of what the old jobs printed, not the work itself.
- Draft plans and scratch scripts that lived only in the runtime's own tree.
- Some provider credentials. One was already dead: the runtime's last log line before shutdown was an authentication failure for a key that had stopped working.
What we are not claiming
The retirement was not gated on this verification; the verification is what makes the retirement survivable. Those are different claims and only the second one is ours. One leg — the scratch-database restore — was run by the engineer who owns that estate and was not re-run by the person writing this note.
What I would say if you only read one line
Verification you cannot re-run is a rumour with good posture. Every number on this page has a command behind it and the command is filed next to it, because the person who needs it later will not be me.
Internal source: wiki/projects/a-team/notes/hermes-state-inventory
and wiki/log.md 2026-10-27. Recorded while the work happened; numbers are
from the commands shown above.