`RETIRED.md`

20 lines
# RETIRED 2026-07-09 — do NOT trust this pre-scan

A 5-critic contrarian panel + a unanimous 3-LLM DTD (verdict B, HOLD) found this
settlement pre-scan to be a **false-green**: it hand-rolled the SHA-locked DW
settlement detectors as a single mega-prompt to a local 7B vision model, returned
0 BLOCK / 0 NEEDS_REVIEW across 76 bird-heavy chinoiserie panels, and on row
SY-112B the model's own prose described a clear violation while the A3 boolean was
biased to false, single-handedly preventing the BLOCK.

**Decision:** approximating a legal-exposure gate with a 7B model is the wrong
instrument. The genuine `settlement` skill (`settlement-post-gen-vision` +
`settlement-verdict`) is the ONLY trusted gate, and it runs at the Shopify DRAFT
import — which is itself Steve-gated. Do not build a parallel approximate
adjudicator.

State: the 8h detached job was killed; all `settlement_verdict`/`settlement_reason`
values were cleared to NULL in fromental_catalog / gracie_catalog / zuber_catalog
(no false data persists). The columns remain (NULL) for the REAL gate to populate
at import time if desired.