This note records the first local CLiD score bridge on the prepared member/nonmember prompt-conditioned split. It is a GPU smoke-scale result, not admitted benchmark evidence.
tiny bridge produced a reusable score schema but is not promotable
Why:
- The local bridge successfully generated member and nonmember score outputs.
- The score-summary schema validator passed.
- The strict-tail metric was nonzero on this tiny split.
- Promotion is blocked because the split has only 8 member and 8 nonmember rows, below the configured minimum of 100 per split.
| Field | Value |
|---|---|
| Run | clid-local-bridge-preflight-20260501-r1 |
| Mode | local-bridge-pair-summary |
| Member rows | 8 |
| Nonmember rows | 8 |
| Best alpha | 0.3 |
| AUC | 1.0 |
| ASR | 1.0 |
| TPR@1%FPR | 1.0 |
| TPR@0.1%FPR | 1.0 |
| Schema status | ready |
| Promotion status | not_eligible |
| Promotion blocker | minimum_sample_gate |
- This does not change admitted evidence.
- This does not change Platform or Runtime schemas.
- This does not establish CLiD as better than
recon. - This only justifies the next bounded packet: collect at least 100 member and 100 nonmember rows under the same score-summary gate.
Run payloads, generated images, raw scores, and schema-review JSON stay ignored
under workspaces/black-box/runs/.