`BLEED_GHOST_TRIAGE.md`

51 lines
# BLEED_GHOST Cohort Triage — read-only

> **Zero DB writes.** Autonomous triage report (2026-06-01). Publishing/unpublishing
> stays **Steve-gated**; this only surfaces what's where. `settlement_verdict` here
> is the bleed/ghost QUALITY verdict, NOT the Part-A/B settlement gate.

## Scope
All `spoon_all_designs` rows with `settlement_verdict LIKE 'BLEED_GHOST%'` — **7,875 designs**.
(The board said "749" — that was stale; the real cohort is far larger.)

## ✅ Compliance check — clean
**0 BLEED_GHOST_BLOCK rows are live.** All **1,600** BLOCK designs are `is_published=false`
AND `user_removed=true` — correctly contained, nothing leaking. No settlement action needed.

## Verdict × state cross-tab
| verdict | published | removed | count | note |
|---------|-----------|---------|------:|------|
| BLEED_GHOST_BLOCK | false | true | 1,600 | ✅ contained — leave as-is |
| BLEED_GHOST_OK | false | true | 2,208 | curator already rejected — leave |
| BLEED_GHOST_OK | false | false | 1,511 | **clean, unpublished — review-for-publish pool** |
| BLEED_GHOST_OK | true | false | 1,333 | live & clean — fine |
| BLEED_GHOST_OK | true | true | 850 | ⚠ **live BUT flagged-removed — contradiction** |
| BLEED_GHOST_REGEN | false | true | 301 | needs regen + already removed — leave |
| BLEED_GHOST_REGEN | false | false | 69 | needs regen, unpublished — regen candidates |
| BLEED_GHOST_REGEN | true | false | 3 | ⚠ live but flagged needs-regen |

## Two things that need a decision (Steve-gated)

### ⚠ 1. 850 designs: `BLEED_GHOST_OK` + `is_published=true` + `user_removed=true`
Live on the site yet marked user-removed — a state contradiction. Either the removal
should take effect (unpublish them) or the removal flag is stale (clear it). **Bulk state
change → needs your call.** Per standing rule, removing live-but-flagged content is
compliance-safe; clearing the removal flag to *keep* them live is the riskier direction.
id list: `/tmp/bg_live_removed.txt` (850 ids).

### ⚠ 2. 3 designs: `BLEED_GHOST_REGEN` + live
REGEN = flagged as needing regeneration, but they're published. Small enough to eyeball.
id list: `/tmp/bg_regen_live.txt` (3 ids).

## The actual "review" pool (the board's ask)
**1,580 clean-but-unpublished** (1,511 OK + 69 REGEN), `user_removed IS NOT TRUE`,
`is_published=false` — the designs that *could* be reviewed for publish. These pass the
bleed/ghost quality check but still need the **settlement gate** before any publish
(BLEED_GHOST_OK is a bleed verdict, not settlement clearance).
id list: `/tmp/bg_clean_unpub.txt` (1,580 ids).

## Recommended next step
1. Decide the 850 contradiction cohort (unpublish to honor removal = safe default).
2. Eyeball the 3 live-REGEN.
3. For the 1,580 review pool: run the settlement gate before any publish — do NOT bulk-publish on the bleed verdict alone.