Complete reviewed evidence inventory
The experimental three-passage representation discarded four supported passages from a reviewed English story solely because they appeared later in its source. The new offline buildActivityEvidenceInventory preserves usable evidence before representation selection. It changes neither production serving nor the existing three-aspect artifact/scorer. More retained evidence is not demonstrated ranking improvement.
Contract and verification
Exact serialized snapshots pass the existing independent eligibility gate before observations are parsed or baseline/source preparation occurs. Observations bind the whole snapshot revision and canonical packet identities. Every expected packet is counted; empty proposals, failed requests and unreviewed packets remain distinct. Unknown/duplicate packets, wrong source hashes, invalid spans and malformed inputs reject with one sanitized error. Bounds are 4 MiB serialized observations, 512 packets, six reviews per packet, and 16 MiB serialized output; overflow is rejected, never silently truncated. These are engineering ceilings, not calibrated quality or capacity settings.
All original validated reviews remain, including unsupported and ambiguous judgments. Conflicting spans cannot produce a supported candidate. Overlapping supported spans prefer fuller context; exact duplicate text in the same scope is deduplicated across sources, while alternative opening scopes stay separate. Stable evidence identities bind source content and offsets. The baseline is byte-identical. There is no first-three or 64-world-review cutoff in the inventory.
Processing coverage is not semantic recall: source collection is bounded, and the model can miss activities or assign wrong verdicts. Output records source omissions, all packet statuses, original observation hash, full revision, eligibility manifest, baseline/source hashes and segmentation runtime.
Thirteen focused tests pass. The initial twelve contract tests failed against the stub before implementation. Independent review then reproduced a duplicate in two global metadata fields; its regression failed before the scope-aware fix and passed afterward. The reviewer independently reran all thirteen tests with no remaining findings. The affected integrated selection passes 577 tests, with zero failures or skips. Workspace typecheck (eight tasks) and build (five tasks) pass. No dependencies or database schema changed.
Complete Discovery CI and OSS export CI passed on source commit fc5598a0b5994530cfe1a6d30b6f2d86660a64a1. The redundant manual Discovery run was cancelled when the delayed PR-triggered run appeared; it is not counted as a completed pass.
Development reconstruction
All twelve prior development snapshots now have a bounded text eligibility decision: three exact snapshots are allowed for offline research; seven families are held after identified exclusions and two remain uncertain. Allows have complete text review, including nested/disabled/UI text. Clear holds can stop early and retain explicit partial coverage. These are AI text judgments, not human moderation, media review, certification or production publication changes.
The current manifest denies all nine held families before preparation, including the final large Chinese snapshot. Only the three allowed snapshots reuse the original immutable model observations. No new model requests, embeddings, rankings or changed semantic verdicts occur.
| Development snapshot | Source fields | Packets | Original reviews | Supported candidates | Old artifact passages |
|---|---|---|---|---|---|
| English team drama | 80 | 3 | 14 | 7 | 3 |
| Japanese story | 61 | 2 | 6 | 1 | 1 |
| English fantasy | 23 | 1 | 4 | 2 | 2 |
All 24 original reviews survive: ten supported, thirteen unsupported and one ambiguous. No conflicts, duplicates or source omissions occurred in these three observations. The English team-drama baseline and old artifact remain exactly equal to the previous reconstruction; its later trust, information-sharing, press-response and draft-access passages are now retained in the inventory.
The Japanese supported passage is disputed by the main review: it promises to continue narration and react to the reader's input, which is generic narrator behavior explicitly excluded by the semantic rubric. Its original supported verdict remains in the audit record; it is not accepted as evidence of distinctive story activity or multilingual quality. The English team example also retains the previously noted over-attribution of an unspecified discoverer to the reader. The fantasy selections describe small opening interactions, which alone do not establish that the representation captures the story's central appeal.
These findings reinforce the need for a separately frozen semantic/selection comparison before spending on a new ranking experiment. The inventory solves information loss, not judge accuracy, salience or the prior failed ranking result. The existing production feed and experimental rollout flags remain unchanged.
Reproducibility
Private-text artifacts stay ignored in the controller workspace. Committed reports contain aggregate evidence and hashes only.
| Artifact | SHA-256 |
|---|---|
| Original twelve-world snapshot | 5f535464d471f7ced86a4ea650b82ef0fe48e5b55306969955587a3f262eb94d |
| Original stopped model run | b5e101f20d7aa43b9edc4c0021e7879bc2d1ab83a29471c379c1da62bb911384 |
| Five additional text screens, v3 | e41bf4b0612d83e4faf89a4df96aed519081811af002966351e2f805688aded9 |
| Final large-snapshot text screen, v4 | 082f45ee01d8cabe66e0fafe85cafb966a3b15edcfa8b6ead4232b123621d516 |
| Current eligibility manifest, approved-v4 | 2f4bddc572267e83953ad840ee8906ef1364daf0298b99e081979c4391992137 |
| Initial inventory reconstruction, replay-v1 | 478945e594aa0bb97e821f35afae7865b91b96043cc0cad7e7d49545db883674 |
| Current-manifest reconstruction, replay-v2 | c3649ca811c6c461b55594711fca0723c9c66fdfbed20d9e2f89964251cc436e |
Replay v1 used eight holds and one unreviewed denial. Replay v2 binds the final nine-hold manifest, verifies unchanged candidates/reviews/baselines for the three allowed snapshots and verifies all nine other families denied. Both preserve the old model run rather than retroactively validating its unscreened protocol. The result records exact helper/dependency hashes and Node/ICU versions.
See previous reconstruction, reviewed snapshot boundary and failed ranking evidence. Frozen evaluation data and the consumed ranking comparison were not accessed for this implementation.