The cleaner deleted the work it was meant to protect.
An agent finished inside an isolated copy. Before its patch reached durable storage, another supervisor treated that live copy as crash debris and removed it.
The answer survived. The handoff did not.
One run on 16 August worked for about 40 minutes and exited normally. Its result was still marked failed: diff capture could not open the patch because the isolation directory had disappeared. A second run completed the same assignment minutes later.
The weekly ledger counted two terminated jobs and two exit-zero failures. It grouped all four as dead from recorded status and explicitly declined to infer beyond those records. This autopsy therefore proves the deletion mechanism and one direct loss. The four-job total is the week’s impact boundary, not four independently traced reproductions of this root cause.
A namespace was mistaken for ownership.
Each write job ran in a disposable clean copy. At supervisor startup, a sweep removed old copies whose names began with the isolation prefix. The design assumed that a shared instance name meant every matching directory belonged to the process doing the sweep.
Instance names can be reused. When two supervisors overlapped under the same name, the newer one could sweep a directory still owned by the older live process. The worker’s prose answer could already exist elsewhere while its uncommitted file changes remained only inside the deleted copy.
Make life visible before deleting.
Changed
- Every isolation copy carries an ownership record with its process identity and running state.
- The sweep checks whether that owner is still alive and leaves live work untouched.
- Deletion writes a durable tombstone naming the reaper.
- Patch capture starts as the first action after the agent exits and streams directly to the outbox.
Still unproven
- The ledger cannot attribute every one of its four deaths to this race.
- One week and one household do not establish a long-run loss rate.
- The fix has a forced race test, but no published multi-week production survival record yet.
- A machine-level failure can still interrupt work before the patch stream begins.
The regression test requires zero removals for a live owner. A dead-owner case requires one removal and checks that the reaper identity survives deletion. Fusion commit 6b0a95c contains the repair and forced test.
Public claim, private operational record.
Counts: private weekly evidence ledger 2026-W33, generated 16 August 2026. Mechanism and regression proof: Fusion commit 6b0a95c. The public page carries mechanism names and aggregate outcomes; operational locations, credentials, and machine-specific details remain private.
Return to the evidence desk · Read autopsy 01: the dead night