`docs/bleed-ghost-settlement-sweep-2026-06-09.md`

107 lines
# BLEED_GHOST settlement-gate sweep — 2026-06-09 (READ-ONLY)

Yolo task `bleed-ghost-review-pool`, Option 1 from the 06-01 brief: text-rule
settlement classification over the bleed-clean review pool. **Zero DB writes**
(session forced `default_transaction_read_only=on`), zero publishes, zero
metered API calls — pure keyword evaluation via `scripts/settlement-gate.js`
`checkPrompt()` (SG-1..SG-4 attorney-narrowed, defendant-favorable, mirrors
`~/.claude/skills/settlement*` Part A1/A2/A3 + Part B + Acceptable carve-outs).

## ⚠ Pool drift — the 1,059 list no longer exists as such

`/tmp/rp_reviewable.txt` is gone and the 06-01 tier-1 definition
(`settlement_verdict='BLEED_GHOST_OK'`, no `EDGES_FAIL` in notes,
`is_published=false`, `user_removed` not true) matches **0 rows today**.
Between 06-01 and 06-09 the pool was dispositioned by curation:

| 06-01 state (triage) | 06-09 state |
|---|---|
| OK + unpub + not-removed: 1,511 (1,059 reviewable + ~464 edges-fail) | 285 (ALL now carry `EDGES_FAIL`) |
| OK + published + not-removed: 1,333 | 2,155 (**~822 published since**) |
| OK + unpub + removed: 2,208 | 2,658 (**~450 rejected since**) |

No disposition timestamps exist, so exact 06-01 membership is unrecoverable.
The sweep therefore covers the two pools that matter now (2,517 designs):

- **`remaining-unpublished` (304)** — every still-unpublished, not-removed
  `BLEED_GHOST_OK` (285, all edges-fail tier) + `BLEED_GHOST_REGEN` (19) row:
  the live descendants of the review pool.
- **`published-retro-audit` (2,213)** — every published `BLEED_GHOST_*` row
  (incl. the ~822 published since 06-01): free compliance canary over what
  already went live through the publish trigger.

## Verdict counts

**Both pools, row-prompt AND lineage-aware: 0 BLOCK, 0 NEEDS_REVIEW, 2,517 OK.**

### remaining-unpublished (304)

| family | OK | NEEDS-REVIEW | BLOCK |
|---|--:|--:|--:|
| stripe | 196 | 0 | 0 |
| chinoiserie | 39 | 0 | 0 |
| designer-zoo-calm | 23 | 0 | 0 |
| muybridge | 14 | 0 | 0 |
| lips | 10 | 0 | 0 |
| drunk-animals (incl. drunk-zoo-36) | 8 | 0 | 0 |
| designer-scenic | 5 | 0 | 0 |
| Traditional | 4 | 0 | 0 |
| archive | 3 | 0 | 0 |
| inspired | 1 | 0 | 0 |
| floral | 1 | 0 | 0 |
| face-skull-damask | 0 | 0 | 0 |
| **TOTAL** | **304** | **0** | **0** |

### published-retro-audit (2,213)

| family | OK | NEEDS-REVIEW | BLOCK |
|---|--:|--:|--:|
| chinoiserie | 519 | 0 | 0 |
| stripe | 416 | 0 | 0 |
| drunk-animals (incl. drunk-zoo-36/monkeys-v2) | 351 | 0 | 0 |
| muybridge | 286 | 0 | 0 |
| designer-zoo-calm | 272 | 0 | 0 |
| monterey-mural | 145 | 0 | 0 |
| designer-scenic | 79 | 0 | 0 |
| inspired | 44 | 0 | 0 |
| cactus + cactus-murals | 56 | 0 | 0 |
| face-skull-damask | 5 | 0 | 0 |
| other (tree-mural/recolor/archive/etc.) | 40 | 0 | 0 |
| **TOTAL** | **2,213** | **0** | **0** |

## Why 0 BLOCK is credible (gate is not dead)

- Control prompt ("tropical banana leaves, directional fronds… open space…
  parrot") → **BLOCK** ✔
- Partial signals across the 2,517: Part B elements hit **54** rows
  (birds/parrots), A2 open-space **1,187**, A3 multi-color **951**, Acceptable
  carve-outs **1,850** — but **A1 directional-foliage fires only 6 times and
  never co-occurs with Part B**. BLOCK requires full Part A AND Part B AND no
  carve-out; the directional-tropical-foliage language was scrubbed from
  generator prompts after the original drunk-animals block wave, and the
  already-BLOCKed cohort (1,331 `BLEED_GHOST_BLOCK` + 475 `BLOCK`, all
  contained unpub/removed) is excluded from these pools by definition.
- 1,828/2,517 rows are fixer-derivatives ("Fix-design of #X…") whose own
  prompt is a pipeline note; the sweep additionally walked parent chains
  (`parent_design_id` + inline `#NNNN` refs, depth ≤6) and took the WORST
  ancestor verdict (`verdict_lineage`) — still 0 BLOCK / 0 NEEDS-REVIEW.

## Caveats

- Text-rule only per task hard limits — **no vision pass**. The 54
  Part-B-present rows (birds/parrots, mostly drunk-zoo/designer-zoo/recolor,
  some published) are OK under the defendant-favorable reading (Part B alone
  is not a violation), but they are the natural first candidates if a
  post-gen vision audit is ever wanted.
- This sweep stamps nothing: `settlement_verdict` in the DB still carries the
  bleed-quality verdicts. Publishing any of the 304 still goes through
  `settlement_publish_check` normally.
- The 285 unpublished OK rows are ALL edges-fail tier now — settlement-clean
  per this sweep, but they need edge-heal before they're publish-reviewable.

## Files

- Per-id verdicts: `~/Projects/wallco-ai/data/bleed-ghost-settlement-sweep-2026-06-09.jsonl` (2,517 lines, ~1.0 MB)
- Script (reusable): `~/Projects/wallco-ai/scripts/settlement-sweep-readonly.js`
- This summary: `~/.claude/yolo-queue/output/bleed-ghost-sweep-2026-06-09.md`