Voice Fact-Lock Extraction
Tags: abbey-root • research • voice-analysis • fact-lock
Voice Fact-Lock Extraction
The fact-locked Voice Model could preserve a reviewed set of propositions, but its upstream locks were still manually authored. This session built the review-gated extraction process for new requests.
Five request types were tested: ordinary post, edit, callback, sensitive message, and factual notice. The local worker proposed fact locks, deterministic validation checked their structure, and machine-readable human reviews drove complete replacement proposals until one was approvable.
The review history was as valuable as the final result. It exposed omitted context, duplicated facts, changed relationships, brittle anchors, confused format constraints, overbroad regexes, and schema regressions. A first approved lock failed during generation, so approval was reopened and the upstream proposal corrected.
The final VOICE-FACT-LOCK-002 passed deterministic generation validation, semantic verification, and human proposition review for all five requests.
The process is now operational but deliberately not automatic. Human review remains the authority for factual completeness and relationship preservation. The next platform improvement should wrap proposal creation and validation in an Abbey command while keeping approval explicit.