tegwick
642163a1ed
Assistant: codex Assistant-Model: gpt-6-astra Assistant-Session: 01a0cffe-206f-7323-89d3-b20ca4cf4bf9
2.8 KiB
Shared guidance revision — 2026-09-26
Status: editorial revision after the user-submitted model-assisted review. This changes the current research corpus, not the frozen pilot. No additional inference sessions, new ratings, human validation or admission are claimed.
| Review gap | Current guidance change | Remaining evidence |
|---|---|---|
| S02: expired idempotency and lost identity | Mutation Outcome Tracking: recover identity, lookup outcome, respect retention; unresolved intent stays uncertain | Actual receiver retention and lookup contracts |
| S03: retention/export assumed after access changes | Local Draft State, Conflict Copy, Explicit Reconciliation: preservation conditional on policy, export separately authorized | Application policy and revocation tests |
| S04: local cancellation confused with remote recovery | Mutation Outcome Tracking: cancellation races, authorized reversal and its own uncertain outcome | Domain cancellation/compensation contract |
| S05: obsolete operation can survive restart | Pending Synchronization: atomically save replacement and retire old pending work; in-flight old requests remain a separate problem | Storage crash and receiver-precondition tests |
| S06: local save/sync overstated as protection | Local Commit Boundary: separate flush claims, independent backup scope and restore test | Executed durability and restore evidence |
S01's request-generation rule remains separate from server freshness: discarding a superseded response does not prove that the retained result is current. Existing VC-01 guidance already distinguishes these obligations. No new pattern is needed for these corrections; the six affected records remain hypotheses.
Evaluation integrity
The pilot's corpus commit remains 78bd95156832c9c48c0e17947a406396c7660088;
preregistration remains 6acbe6d. Packs, prompts, assignments, raw receipts, original
scores and submitted adjudication are unchanged. Frozen input hashes must now be
checked against those historical Git objects, not substituted with current drafts.
The old runner's current-tree validation is expected to reject revised corpus files.
Reproduce the old study from its frozen checkout; do not refresh its manifest.
The six scenarios are consumed development examples. A later evaluation must freeze the revised corpus, use new scenarios and score atomic obligations. Its purpose would be to test improvement, which these editorial repairs alone cannot establish. WP-0002 T03 retains the independent-human evidence gate.