27 lines
5.1 KiB
JSON
27 lines
5.1 KiB
JSON
{
|
|
"candidate": "9641ed52def2ff8e5b0222449a4a83a1191e1d37",
|
|
"scope": "Limited new preservation/full-restored-snapshot scenario evidence review; not formal whole-unit acceptance",
|
|
"result": {
|
|
"status": "changes_requested",
|
|
"summary": "Limited review of one native macOS producer→consumer run on 9641ed5. Adequately shown: different processes (PIDs 39701 and 40866, recorded at launch and per observer row), the same recorder hash in both runs, and the controller guards (exact commit, clean tree, source/stage/default-player unchanged, exit 0, no ERROR in the log). Full restore equality at snapshot_published before the ordinary SAVE_LOADED event is also adequately shown: the snapshot the observer built at the publish and the snapshot read back from disk are byte-identical (SHA b242cc87…, 118528 bytes, both pinned), the only capture was index 0, and a fresh process can only match the minute-444 save if the restore really happened. This holds for that phase only, not after the event. NOT adequately shown: preservation of ALL FOUR old slot files. Only manifest.json and world.bin are re-hashed on the live manual_ui path. journal.log and thumbnail.png rely only on preexisting-slot-before/after.json, which neither shown controller nor the observer writes. Their method, author and timing relative to the named-slot save are not shown, and their size difference (1472 vs 762 bytes) is unexplained. Menu ids 3/2 and the resume results (clock 444→531, reception finished, gold 4500/material 42) are CUA observations only, with no pinned captures. The restore-equality and different-process parts would pass alone; the ALL4 claim does not.",
|
|
"findings": [
|
|
"PASS-LIMITED: Restore equality is shown at the right phase. The observer calls runtime.create_snapshot() when snapshot_published fires. When the recovery completes, it compares that snapshot with the disk read from inspect_snapshot using == and compares SHA-256 of both serialized byte forms. snapshot-000.bin and loaded-stored-snapshot.bin are pinned with identical SHA b242cc87… / 118528 bytes. snapshot_index 0 and a fresh process mean equality is only possible after the restore. Valid only before the SAVE_LOADED completion event, as stated; not post-event equivalence.",
|
|
"PASS-LIMITED: Different processes. PIDs 39701 and 40866 come from each controller's Popen; the observer also logs its own process id per row. The consumer reused the producer's profile and required it to exist.",
|
|
"PASS-LIMITED: Guards. Each controller checks the exact candidate commit, a clean tree, tested stage source hashes, and no other running engine. Afterwards it checks source refs, the stage inventory and the default player are unchanged and restores project.godot. A run passes only with exit 0 and no ERROR/SCRIPT ERROR in the native log. ended.json is hash-pinned, but its contents were not supplied, so passed=true and empty guard_errors are taken from the summary.",
|
|
"GAP (blocking ALL4): Only manifest.json (62d801bb…) and world.bin (da65a181…) are re-hashed on the live saves/manual_ui path. journal.log (1ba101ad…) and thumbnail.png (17b547b7…) have no live re-hash pins and rely only on preexisting-slot-before/after.json.",
|
|
"GAP (blocking ALL4): Neither the producer nor the consumer controller, nor passive_os.gd, writes preexisting-slot-before.json or preexisting-slot-after.json. Their author, method and timing are unproven. The before capture must fall after the quick save and before the named-slot (menu id3) save. Its 1472 bytes vs the after file's 762 bytes is unexplained. Supply both files' contents and the capture method, with timing tied to the events.jsonl ticks.",
|
|
"SCOPE NOTE: The producer required a fresh, empty profile (require not profile.exists(); mkdir). So the 'preexisting' manual_ui slot was created by the same run's quick save. It existed before the named save, not across sessions.",
|
|
"OBSERVER LIMIT: Menu ids id3/id2, the 'preservation-r03' name entry and the confirm-read are not proven by pinned data. The observer logs _input keys and mouse buttons and dialog/label state (contents not supplied), but not PopupMenu item ids. It may not see clicks inside popup subwindows. These rest on CUA observation.",
|
|
"OBSERVER LIMIT: Resume results (clock 444→531, reception 70%→completed, gold 4500/material 42) are CUA-only. Only snapshot-000 is pinned; later snapshot_published captures were not supplied.",
|
|
"MINOR: The observer is not perfectly passive. Connecting to snapshot_published makes _publish_snapshot emit the signal, which it otherwise skips, and the handler adds a create_snapshot() call. This is likely harmless for state but does change the code path.",
|
|
"NOT CLAIMED/NOT ACCEPTED: whole P24/P24-02, all content, Windows/other platforms, post-event equivalence, repeated full roster."
|
|
],
|
|
"whole_unit_accepted": false
|
|
},
|
|
"artifact": "/Users/korizatou114/Library/Application Support/agent-control/the-tree-p24-02-operations/runs/3b4f57954a9d44b299c47feb4b024139",
|
|
"model": "claude-opus-5-5",
|
|
"formal_runner_review": false,
|
|
"whole_unit_accepted": false,
|
|
"input_proof_sha256": "024a2cb0aba292037f281a3eea43b5eac95d773fa0517e77b32b08ef5084e7c1"
|
|
}
|