25 lines
3.7 KiB
Plaintext
25 lines
3.7 KiB
Plaintext
Perform the actual independent DIRECT IMAGE review for P23-03 on frozen candidate 49445d4.
|
|
This is auxiliary image evidence production, not normal all-condition Runner.review and not acceptance.
|
|
Read the latest .review-material/policy/OPUS_R13_VISUAL_ORDER_DIRECTION.json, which permits this VISUAL-ONLY pass before full177,
|
|
then the earlier R16 proof-method recipe and immutable docs/spec/acceptance/P23-03.json. Source/spec/gates are unchanged.
|
|
Complete original R13 proof is mirrored in .review-material/proof. Regression-r01 stopped on disk admission after58 PASS;
|
|
its unchanged incomplete receipts are in .review-material/regression-r01-incomplete. One fresh complete full177 remains mandatory
|
|
before normal validation/all-condition review/integration/final/ledger. Do not award any regression or unit credit in this pass.
|
|
Original absolute proof paths in raw receipts map to .review-material/proof; raw bytes/SHAs were preserved, not relabelled.
|
|
Use .review-material/VISUAL_ROWS.json as the complete eligible row index; original manifests/crops and whole screenshots are available at its paths.
|
|
For EVERY manifest and EVERY eligible row, directly inspect the rendered category, expected state/operation/text/dimming and position against the source-grounded observation.
|
|
Open the actual whole screenshots and necessary crops with Read. Inspect relevant world_preview.gd/room_presentation.gd and source state enums as needed.
|
|
RGB fields are provenance, not an independent classification oracle. No invented classifier/noise threshold is permitted.
|
|
Read any additional original proof or source you need: there is no fixed file/byte cap. CROP_PIXEL_EQUIVALENCE.json indexes exactly byte-identical crop files only;
|
|
you may read identical pixels once, but evaluate each row's expected state, position and context independently. Do not infer a pass from that index.
|
|
Report false for any mismatch or required image/observation you cannot verify, with concise findings identifying manifest/row. Worker/Root claims are not judgments.
|
|
Output reviewer='independent-opus-direct-image', all frame manifest_sha256 values and rows in the original eligible row_index order.
|
|
For compact lossless serialization, each row is EXACTLY [row_index, category_boolean, state_boolean, position_boolean].
|
|
All three booleans must be your explicit independent decisions. Root will only rename the four tuple positions into checker keys without changing decisions.
|
|
Do not omit a manifest or row, auto-fill repeated labels, write verdict files, edit sources, run tests, dispatch workers, or modify task state.
|
|
Return only the requested structured result, plus concise summary/findings fields; no long per-row prose.
|
|
|
|
Latest actual Opus batching direction: all requirements remain, but THIS CALL outputs ONLY the frames in .review-material/serial-image-workload/batch-01.json. Read that index first.
|
|
The actual recorded commander direction is .review-material/serial-image-workload/OPUS_R13_VISUAL_BATCH_DIRECTION.json. Its serial partition replaces the earlier bulk output assignment, without changing the meaning of any judgment.
|
|
Return all assigned frames in the index order, with every eligible row in its index order, including empty row arrays. You retain access to ALL original materials and sources. The output workload split is not a file/byte reading cap. Do not read the entire 1780-row index merely to find this assignment; the assigned index faithfully copies the complete assigned rows and paths. Read actual assigned images/crops and needed source. Exact identical pixels may be read once, but assess each row context and explicitly output its three booleans. Do not produce decisions for other batches. Prior bulk timeout supplied zero decisions. No source edit, test, subagent or task-state change.
|