Bounded Facebook Voice Model
Tags: abbey-root • research • voice-analysis • voice-model
Bounded Facebook Voice Model
The first Voice Model now exists.
It synthesizes four validated Facebook characteristics: deadpan contrast, semantic compression, selective continuity, and quoted stance marking.
The model is deliberately bounded. It does not claim to describe technical, professional, academic, long-form, or sensitive writing. The interactions between characteristics are provisional because the research validated the components, not every combination.
The application evaluation used eight invented scenarios and did not provide the Facebook corpus to the worker. That made the test about abstraction rather than retrieval or copying.
The worker did several things well. It avoided humor in a family-illness message, kept a technical notice plain, used supplied callback context without inventing earlier history, and did not mistake an ordinary album title for quoted stance.
But the first run invented a maintenance time and changed the factual premise of an editing task. The model was revised to prohibit invented facts and require factual preservation.
The second run failed in the same general way with different details. It invented midnight, changed a family member into an aunt, added a fifteen-minute restart schedule, and replaced a toaster failure with a firmware-update premise.
Both outputs looked plausible. Neither was accepted.
That is an important result. The research model can describe validated writing characteristics, but a free-generation prompt is not a safe application workflow. Stronger wording in the prompt did not reliably preserve supplied facts.
VOICE-MODEL-001 remains a bounded research draft. The next application experiment should lock factual propositions separately from stylistic transformation and reject unsupported changes deterministically.