Fact-Locked Voice Application
Tags: abbey-root • research • voice-analysis • voice-model
Fact-Locked Voice Application
The two failed free-generation evaluations established that stronger prompt language was not enough to preserve supplied facts. The next application method therefore separated immutable propositions from stylistic choices.
A machine-readable fact lock now defines required propositions, protected literals, authorized creative slots, characteristic boundaries, format requirements, and known unsupported additions for the eight evaluation scenarios. The generated output must pass deterministic checks before it is eligible for a separate semantic review and human scoring.
This layered workflow caught several important problems during development: partial JSON, a creative-slot schema mismatch, a tense change from an ongoing action to a completed action, and invented device agency. It also demonstrated that the semantic verifier is advisory because it initially missed the tense change.
The final gpt-oss:20b candidate passed all eight deterministic checks, all
eight semantic-verification items, human proposition review, and the original
application rubric with a score of 78 out of 80.
VOICE-MODEL-001 version 3 is now approved only through this fact-locked workflow and only within the tested Facebook scope. Free generation remains unapproved.
The next unresolved platform problem is upstream: deriving a trustworthy, reviewable fact lock from a new writing request before the style model runs.