MENS curriculum (speech-to-code)
Staged supervision to reduce “lost in transcription” drift:
- Stage A — Transcript cleanup:
asr_refineand deterministic Oratio refine pairs; teach model to fix ASR noise without changing CLI flags/paths. - Stage B — Intent / structure: Short prompts mapping normalized transcript → outlines (function names, parameters) without full program.
- Stage C — Constrained codegen: Full
.voxemits with compiler-checked examples only (speech_to_codemix rows). - Stage D — Repair supervision: Prompt = failing snippet + diagnostics; response = minimal fix (MCP retry-loop style).
Weight higher-quality, compiler-validated rows; cap aggressive ASR-only pairs. See speech-to-code-pipeline.md and mens-training.md.
QA / labeling
Use contracts/speech-to-code/labeling_rubric.md for human or LLM-assisted labels (intent_ok, compile_ok, semantic_ok, verbatim-sensitive spans). Export traces with failure_category (not a loose free-form category string) for KPI joins.