Field note · October 17, 2026
Sixteen GiB.
Half a GiB back.
Eight temporary Chrome signing copies were archived, verified, and removed. The disk gained about half a GiB. A companion comparison stayed paused.
The comparison reached the machine.
A fresh Luma comparison had moved from a generation-only replay to the companion’s actual app route, with normal memory available. Four requests returned. The fifth timed out. The remaining seven requests were not submitted.
During the investigation, a local write failed with “no space left on device.” The API also stopped answering health checks. Those events occurred together; the evidence did not establish one exclusive cause of the timeout.
The big number was a candidate.
A read-only inspection found 23 temporary Chrome code-signing copies reporting roughly 47.5 GiB in directory usage. That was a place to investigate, not a promise of recoverable space. Preserved conversation sources were excluded from cleanup.
The repair was deliberately bounded: choose the eight oldest inactive copies, archive them to a separate external volume, verify them, and request a separate confirmation before removing their internal originals.
The archive contained about 16.01 GiB of regular-file content. Relative paths, file modes, symbolic-link targets, and SHA-256 hashes of every regular file matched. The copies were revalidated before removal. Only the eight approved internal directories were removed; their external archives remained intact.
The number that mattered came afterward.
The measured increase in free space was approximately 0.50 GiB. The final check reported 4.87 GiB free, below the agreed 20 GiB operating target.
This is consistent with APFS sharing blocks among copies: a directory’s reported size is not the amount of storage exclusively owned by it. Per-block ownership was not measured, and concurrent disk activity means the observed change cannot be attributed byte for byte to these removals.
The pause was part of the result.
The API was healthy at the final check. Luma’s implementation, prompts, model configuration, and personal memory were not changed by the disk repair. The timed-out diary request was checked and had no persisted response.
The comparison remained paused because the storage target had not been met. The failed attempt stayed in the record. No dialogue was reconstructed, no request was silently retried, and no broader cleanup was substituted for the approved eight-directory scope.
Infrastructure recovery and evaluation completion are separate things to prove. This note records the first bounded repair and its limit, not a completed Luma comparison.
The public receipt.
Save the measured outcome ↗ Source and publication boundaries ↗