# Working Leo Strict Semantic Rescore

## Result

The post-concision live VPS transcript was rescored without another model call
after adding semantic gates for proposal readiness, current `public.beliefs`,
identity rows, row-content proof, bounded source intake, and a 220-word default.

- strict result: `2/15`;
- passing: `OOS-09` participant identity and `OOS-13` handler-versus-Telegram proof;
- previous phrase-oriented score on the same transcript: `5/15`;
- database counts changed: `false`;
- service unchanged: `true`;
- temporary profile removed: `true`;
- posted to Telegram: `false`.

## Meaning

The stricter result matches the independent review: concision alone did not
make Leo a reliable current-schema or apply-readiness authority. The repair must
change the live readback and doctrine, not merely tune expected phrases.

The two highest-impact gaps are now concrete:

1. approved legacy proposals need strict payload normalization and renewed
   review before any apply action can be offered;
2. shared factual claims and agent-owned positions must use the current
   `public.beliefs` model, whose schema has no direct claim-ID foreign key.