Retrospective entry prepared by Codex; work date 2026-10-29; original draft prepared 2026-11-05T19:20:07.192Z; publication edition prepared 2026-11-05T19:32:31.997Z; historical work and technical-review statements scoped to the retained receipts; publication approved by Asif; first published 2026-11-05; owner editorial review of individual technical claims is not recorded. Publication preparation did not rerun engineering operations or change private runtime.
A real local-model task completed through the desktop harness, but its answer was not good enough to use. The single 3B attempt produced a complete 1,429-byte operating checklist. It repeated the supplied roles, then proposed device synchronization and resource-reclaim behavior that had not been established. The useful result was the separation that followed: successful execution and intact storage did not turn those recommendations into validated guidance.
A separate agent review rejected the unsupported recommendations and supplied a clearly attributed correction. The desktop panel showed “Completed · review not accepted,” with the original model answer and reviewer correction side by side. The original model, attempt and result receipts stayed immutable. The correction was not passed off as something the model had produced, and the review was not another generation.
Recovery was tested through a normal full application quit and relaunch, rather than a page refresh. Reopening the original task recovered the same request, one attempt, result and saved review. The complete journal hash, result bytes and review bytes remained unchanged; the ledger did not grow and no extra model call was made. That established this application-recovery path. It did not prove behavior under a physical network disconnect, reboot or arbitrary device handoff.
The integration also failed in a concrete way before it worked: the first review-storage namespace used a name the native store rejected. No record was written. The name was corrected while the app was closed, and the rule became a regression check. Thirty-nine focused tests covered result and attempt binding, stale responses, conflicting writes and storage failure; 64 render checks complemented the later native storage and restart observations.
The phone-facing work kept the same distinction between evidence and capability. It exposed one specific completed task and its separate review, then added a read-only resource view over four saved inputs. Declared resources, recorded observations and stale caches were labeled separately. The view did not probe devices, allocate work or infer current capacity. Its 32-test Linux check left all input bytes and modification times unchanged.
The remaining gaps stayed visible. Historical running and uncertain jobs were preserved without replay. Broader conversation synchronization and shared allocation were still unfinished, and the mobile layout check remained an offline touch fixture rather than final physical-phone acceptance. The next useful demonstration would need an existing project source, a cited passage, an independently review-passing result and recovery of the same conversation on both interfaces—not another green completion badge.
What did not ship
- The complete model answer did not pass content review; the reviewer correction was not credited to the model.
- Application quit/relaunch did not establish reboot, physical network-loss or arbitrary device-handoff recovery.
- The cached resource view did not allocate work, probe devices or establish final physical-phone acceptance.
Next
Select one existing permitted project source, retrieve a cited passage into one durable task, accept the result independently, then reopen the same conversation/result/review on desktop and phone without creating a second job.
Raw telemetry
- Original model task: one attempt; complete 1,429-byte result; content review rejected unsupported operating advice.
- Full application recovery preserved the same task, attempt, result and review; the ledger did not grow and no new model call was made.
- The retained integration record reports 39 focused tests and 64 render checks; the separate cached resource-view run passed 32 tests with four saved inputs unchanged.