# Class 10 completion diagnostic — no publication

## Outcome

The unchanged multiscale method cannot fill the remaining quota from these frozen inputs.

| Scenario | Existing Class 10 | Explicit existing-human reuse | New multiscale drafts | Attainable | Still missing |
|---|---:|---:|---:|---:|---:|
| Original quota base | 31 | 0 | 3 | 34 | 16 |
| With six validated existing human masks | 31 | 6 | 3 | 40 | 10 |

All nine additions are globally distinct from the original 231 frames and one another. The six reused labels are **existing saved human masks**, not generated masks, not six new human edits, and not transfers. Their complete original current payloads, human-event hashes, full per-target event histories, and formal/classification/source identities are in `existing_human_reuse.jsonl`. They remain unconfirmed drafts. Nothing has been published, imported into the gallery, or written to a live database/configuration.

## Frozen human seeds and protections

- Exactly 175 `quota_generation:<target_key>` templates from the quota inventory: 78 original pose-human + 97 gallery-human; 22 are Class 10.
- Recovered every source from the frozen SQLite current records and actual human events; exact inventory comparison and native gallery-image pixel comparison passed. Pose human-event hashes use original UTF-8 payload bytes, not reserialized JSON.
- Latest live pose/gallery logical hashes matched the quota snapshots before and after generation. The frozen pose snapshot SHA-256 is `7c24e787f5267cde35f73afcb29f5f19c4e4499b83ef23a7b2278dac624aba9e`.
- All 231 prior/manual/new-machine targets excluded. All non-original-mask history, gallery history, protected states, quarantine and human seed targets protected. Original-rectangle-only pose/ruler saves retain the original sidecar eligibility policy.
- Neither the 67 newly selected machine drafts nor the 43 old unreviewed machine entries were promoted to seeds.
- All existing quota artifacts and protected formal hashes were unchanged at generator completion. Separate independent verification of the quota batch remains the coordinating session's responsibility.

## Exhaustive rejection funnel

Only Class 10 targets were scanned; **every compatible template was attempted**, even after a target had already passed. No threshold, window configuration, cross-window requirement or boundary guard was changed.

| Stage: at least one compatible pair/configuration passes | Targets |
|---|---:|
| Eligible remaining Class 10 targets | 412 |
| At least one compatible human template | 213 |
| Initial NCC >= 0.55 | 90 |
| Valid affine registration/deformation | 83 |
| ECC >= 0.84 and boundary NCC >= 0.78 | 38 |
| At least one alternative config passes original transfer + beam gates | 19 |
| At least one alternative config also passes every-vertex local image/gradient support | 8 |
| Two different windows independently admissible and within 1.5 patch pixels at every vertex | **3** |

The 199 targets with no compatible seed split into: 132 without a same-split/same-policy seed, 62 without any same-split seed, and 5 with incompatible known viewing faces. These restrictions were not bypassed.

497 compatible pairs were evaluated: 325 initial-similarity failures, 12 excessive-affine failures, 100 ECC/boundary-correspondence failures, 57 without a fully supported cross-window configuration, and 3 successes. All failure metrics/configurations are retained in `registration_attempts.jsonl`; these are not accuracy probabilities.

Five targets had an individually admissible locally supported alternative configuration but no acceptable distinct-window consensus: `463be349b2cd9a13b555f0c6`, `6890d57090773e451d8c0060`, `9ed65e6b506ba09c29b50160`, `cb5a18714bac548b5060529b`, `e5e794ff419bc90502b5bf34`. They were **not** accepted. Thus photometric compatibility alone does not imply 13 or 19 valid new frames.

## Accepted candidates and quality caveats

| Target | Direct human source | Vertices | Boundary NCC | Max FB patch px | Native refinement / limit px |
|---|---|---:|---:|---:|---:|
| `052f61a64597b0f28c6c4d3e` | `858d34ea2b902a43f28e22c6` | 8 | 0.937690 | 0.089659 | 1.674900 / 5.213245 |
| `0d707a13265e2f6ce6c24f85` | `5646b257c8d3550f6d8d98b2` | 12 | 0.790736 | 0.933650 | 4.660220 / 4.755598 |
| `a563476c61926f13035f3e2a` | `82570f9f6b30bb2c1d696306` | 8 | 0.824040 | 1.039390 | 1.144927 / 5.293428 |

All three selected window 25 / pyramid level 0 and have independently admissible window-31 support. No coordinate averaging, topology resampling, original-method success claim or recursive propagation was used.

Visual inspection of the three generated comparison cards found no reason to manufacture extra successes. Specific review priorities:

- `0d707a13265e2f6ce6c24f85` is close to two strict limits: boundary NCC only 0.010736 above 0.78, and refinement only 0.095378 native pixels below its limit. Inspect the top beam and narrow central support carefully.
- `a563476c61926f13035f3e2a` and its source show sparse/noisy intensity and a strongly slanted beam; edge placement and inherited 8-vertex topology require human judgment despite all numerical gates passing.
- `052f61a64597b0f28c6c4d3e` is visually better supported, but gate/replay success still does not constitute human confirmation or ground truth.

No post-hoc visual rejection or acceptance override was applied. All three remain research candidates pending independent validation **and** human review. No multiscale module bug was found or edited.

## Reproducibility, not independent publication verification

- New driver: `scripts/diagnose_class10_completion.py`.
- Method: `bounded_lk_cross_window_consensus_research_v1`, a separate alternative algorithm.
- Multiscale module SHA-256: `a68efecacf83ca24f46ec2d754752f739680774de137ede073ae33c81052f661`.
- Externally supplied frozen-input SHA-256: `03a91b06988e7063aa432e1f19333cd1db373cc3cdefdc76cd099c5f696f77f9`.
- A fresh-process replay compared all 497 complete pair records, all metric leaves, polygons, 412 target outcomes, source provenance, six human reuse entries and quota selection **exactly**. Independent inequality and fixed-priority consensus checks also passed. See `replay.json`.
- Existing multiscale regression tests: **4 passed**. This is not an independent publication seal.

Replay command (the program deliberately refuses to overwrite an existing replay result):

```bash
.venv/bin/python scripts/diagnose_class10_completion.py --workers 4 \
  --replay-frozen-sha256 03a91b06988e7063aa432e1f19333cd1db373cc3cdefdc76cd099c5f696f77f9
```

`accepted_candidates.jsonl` and `selected_candidates.jsonl` each contain the same three distinct-frame research drafts. `selected_without_human_reuse.jsonl` separately records the original-base scenario. `report.html` and `cases/` show their image comparisons. `frozen_inputs.json`, `frozen_code/`, `seed_source_records.jsonl`, the SQLite copies and the full audit preserve replayable provenance. `artifact_hashes.json` covers generation artifacts; `handoff_hashes.json` additionally covers the subsequent replay and this handoff.
