← Back to Journal

Fact-Locked Voice Application CLI

Abbey Root now has the first public entry point for applying VOICE-MODEL-001: abbey research voice apply. The command refuses unapproved inputs, uses the canonical fact-locked application prompt, and gates its structured output through deterministic validation.

The workflow remains deliberately narrow: it is limited to the reviewed Facebook scope, and semantic verification and human review remain required after deterministic validation.

Smoke-test correction

The first real smoke test correctly failed deterministic validation: three responses used propositionally related wording but omitted required lexical anchors, and the model returned fenced JSON with the model field set to VOICE-MODEL-001. The workflow now explicitly instructs required-anchor coverage, normalizes fenced JSON to standalone JSON, and overwrites the model metadata with the invoked runtime model. The original failed smoke-test output and report remain preserved as diagnostic evidence. A corrected Ollama smoke test was attempted with new filenames but the configured service did not complete before the run was interrupted; no candidate regeneration or fact-lock changes were made.

Second smoke-test finding

The second smoke test passed four of five scenarios. REQ-002 exposed a narrow serialization artifact: parsed response prose contained literal backslash-plus-quote characters, preventing the approved marketed as "smart" anchor from matching. The normalizer now removes only that artifact from response fields, preserves unrelated backslashes, and retains valid JSON serialization. A fresh corrected-003 smoke test was attempted without overwriting earlier diagnostics, but the configured Ollama service did not complete before timeout/interruption.

Deterministic replay validation

The preserved corrected-002 smoke output was replayed through normalization without invoking Ollama. A new replay artifact recorded gpt-oss:20b, removed only the visible quote-escape artifact, and passed deterministic validation for all five scenarios. This confirms the correction independently of fresh generation.

First-real-use prompt generalization

The first real use exposed evaluation-specific prompt language: a fixed eight-scenario requirement, EVAL example IDs, and a Project Lantern-specific slot instruction. The canonical prompt now copies the supplied fact-lock and scenario IDs, preserves source order for any scenario count, and applies creative-slot rules generically. The approved lock, review chain, and deterministic validator were unchanged. Mocked regression coverage confirms a custom one-scenario lock reaches input validation, normalization, and deterministic validation.

First-real-use typography replay

The first one-scenario generation preserved the propositions but substituted U+00A0 NO-BREAK SPACE and U+2011 NON-BREAKING HYPHEN inside protected literals. The normalizer now maps only those two characters in human-facing responses to ASCII space and hyphen-minus. A replay of the preserved failed output, written to new files without Ollama, passed deterministic validation for REQ-001 and retained the runtime model gpt-oss:20b; this was a deterministic replay, not fresh generation.

Human review of first real voice application

Brad reviewed the normalized first-real-use candidate and the supporting diagnostic artifacts.

The approved fact lock is not weakened or retroactively altered. With only four validated voice observations, this limitation is expected. The public fact-locked application workflow is operational, while VOICE-MODEL-001 remains an early bounded model requiring more voice evidence and evaluation. The generated candidate and all diagnostic artifacts are preserved; no replacement candidate was generated.