# Publication 6 generation handoff (not independently verified)

- Output: `outputs/pallet_mask_quota_completion_v1/`.
- State: `quota_pending_independent_verification`; no configuration/service publication changes were made.
- Strict original `attempt_target` only, four CPU workers, unchanged transfer/boundary code and thresholds.
- 4,189 eligible targets scanned; 144 accepted before distinct-frame/quota selection; 67 machine drafts selected.
- Cumulative 231 globally distinct frames = 140 unchanged parent records + 24 explicitly requested saved-human masks + 67 machine drafts.

| Class | Old total | Added saved human | New machine | Total | Missing |
| --- | ---: | ---: | ---: | ---: | ---: |
| 1 | 48 | 2 | 0 | 50 | 0 |
| 2 | 5 | 8 | 37 | 50 | 0 |
| 9 | 38 | 3 | 9 | 50 | 0 |
| 10 | 10 | 8 | 13 | 31 | 19 |
| 15 | 39 | 3 | 8 | 50 | 0 |

## Stable files and schemas

- `proposals.jsonl`: cumulative heterogeneous records; preserve the 140 prior records canonically.
- `new_proposals.jsonl`: **only 67 machine drafts**, schema `direct_human_union_quota_v1`. Use only this file for the adoption sidecar; human sources must not self-transfer.
- `manual_proposals.jsonl`: exactly the 24 completed request targets, schema `saved_pose_mask_gallery_entry_v1`, origin `original_pose_human_revision`, `source_label_kind=human_mask_source_manual_template_draft`, status `human_edited_draft_pending_confirmation`.
- Each manual record contains the current saved pose polygon verbatim, `pose_snapshot_saved=true`, `human_edited=true`, `metrics={}`, and `manual_source_not_machine_transfer`. There is no `production_saved=false` claim. `pose_store_written` and `gallery_store_written` are false.
- Manual template metadata uses its **own target key**, actual saved revision/hash and human polygon event revision/hash, with `pose_review` provenance. `rendering_id` and `template_inventory.json` mapping use `quota_generation:<own-key>`.
- `manual_source_records.jsonl`: schema `saved_pose_mask_source_freeze_v1`; freezes parsed and exact raw current/event payloads. `current_payload_sha256` hashes canonical JSON, but `human_event_sha256` hashes **the exact UTF-8 event payload bytes** used by the original `annotation_inventory`. Do not canonicalize the event when recomputing that hash.
- `requests.jsonl`: byte-identical frozen copy of the original 24-item queue, not a claim that its old pending statuses are current. Completion is represented by the new manual records.
- `reviews_snapshot.sqlite3`: 4,758 saved pose records / 10,290 events; `gallery_reviews_snapshot.sqlite3`: 97 saved gallery records / 104 events. Read-only snapshots and backups retained.
- `template_inventory.json`: 175 direct human generation templates (78 original pose + 97 gallery). Two other pose edits were found beyond the expected original 52 plus requested 24; those, like all other non-requested pose seeds, do **not** add quota entries.
- Class counters explicitly separate `old_total`, `manual_added`/`added_saved_human`, `saved_gallery_human`, `saved_pose_human`, `saved_human`, `prior_machine_pending`, `new_candidates`, `total`, and `missing`. Saved human = 121; prior pending machine = 43.

## Review integration contract

Manual cards/assets show only their frozen saved-human mask with honest labels, no fabricated registration score or machine comparison. Independent gallery review history was **not** populated from pose history. A manual-source detail should initialize from `proposal.polygon`, retain a null independent saved-human overlay until an actual gallery edit, and expose the original pose link/provenance. A frozen source mismatch must be handled as a stale source, not silently imported or confirmed.

All new/manual records remain unconfirmed, ineligible for automatic training, GT or robot commands. The parent 140 objects and 894 historical target/template assets passed canonical/hash preservation checks. Both live databases' full logical tables still matched their frozen snapshots at generator self-check time. Formal protected hashes and frozen algorithms were unchanged.

## Tests and verification boundary

113 targeted/regression tests passed, including 57 quota tests. `generation_validation.json` records generator self-checks only; it is **not** an independent publication verification. Independent verification and explicit publication remain with the coordinating session. Class 10 still lacks 19 frames; do not relax gates or claim quota completion.
