core/docs/decisions/ADR-0183-lawful-audio-lexeme-path.md

ADR-0183: Lawful Audio→Lexeme Path (stub)

Status: Proposed (stub — placeholder to record the fork; not yet a full design) Date: 2026-05-29 Authors: Joshua M. Shay, Core R&D Engine Domains: sensorium/audio/, language_packs/, (future) generate/ Depends on: ADR-0013 (Sensorium Protocol), ADR-0181 (Audio Compiler), ADR-0180 (Delta-CRDT substrate) Related: docs/audio_pipeline_overview.md §9 (teacher boundary), docs/plans/audio-compiler-eval-plan.md §4 (teacher policy)

---

1. Why this ADR exists (context)

The audio compiler (ADR-0181) decodes the paralinguistic layer of speech lawfully and deterministically: energy, voicing, pitch contour (prosody), pauses, turn boundaries, emphasis, non-speech events. It is deaf to lexical content — it hears how something is said, never what words were said.

ADR-0181 admits learned ASR (Whisper/NeMo) as teacher/shadow lanes only — typed transcript labels, never substrate, never folded into the versor. The serving path is and must remain Whisper-free (overview §9): the teacher teaches; the lawful path serves.

That doctrine only holds in production if a lawful runtime path exists to carry the words the teacher bootstrapped. This ADR is the placeholder for designing that path. It is a stub: it records the fork and the constraints so whoever reaches the words problem does not silently reach for Whisper in the serving path. It is not a committed design yet.

This is the PR to scrutinise hard (per the overview's warning): the moment a real consumer turns audio into words at serving time, the 0-parameter, decode-not-borrow doctrine is on the line.

2. Problem statement

Define a deterministic, 0-parameter, replayable path from a canonical audio signal to lexical content (words / lexemes) that:

• never calls a learned model in the serving path;
• preserves the ADR-0181 invariants (versor condition < 1e-6; merge-key determinism; no PCM in traces; no normalization outside allowed sites);
• recovers what it can from physical facts in the wave (decode, not borrow), and refuses honestly where it cannot (no fabricated words — the wrong = 0 discipline generalises here);
• can be taught (vocabulary grown through the reviewed corridor) without the teacher becoming a runtime dependency.

3. Candidate directions (to be evaluated, not yet decided)

Two non-exclusive paths, named in the overview:

• (A) Words-as-text. Audio stays a paralinguistic-plus-coarse-phonetic sense; the what arrives through the existing text modality. No serving-time ASR at all. Cleanest and lowest-risk; may be sufficient for many use cases.
• (B) Deterministic audio→lexeme decode. Extend the acoustic front-end toward lawful phonetics — formant tracking, spectral-band energies — to recover vowel quality and broad consonant classes, then match against taught vocabulary (acoustic-pattern ↔ lexeme associations grown via the reviewed teaching corridor). 0-parameter, fully replayable; real-speech ceiling is lower than learned ASR, and that cost must be stated honestly.

A teacher (Whisper, reviewed) may bootstrap the taught vocabulary for (B), then be removed — it is scaffolding, not a component (overview §9).

4. Constraints any accepted design must honour

• 0 learned parameters in the serving path (track in docs/model_dependency_size_tally.md).
• Decode, not borrow — features must be physically present in the wave; no opaque latents (ADR-0181 §5 rejects embeddings as substrate).
• Refuse over fabricate — where phonetic evidence is insufficient, emit no lexeme rather than a guess; reuse the wrong = 0 / honest-refusal discipline.
• Lexeme growth is reviewed — new acoustic↔lexeme associations enter only through the contemplation → proposal → HITL corridor (cf. ADR-0164/0165's treatment of lexicon/primitive growth), never as raw model output.
• Determinism / replay — same canonical bytes ⇒ same lexemes ⇒ same trace hash; quantize before semantics (ADR-0181 spec §7).

5. Open questions (for the full ADR)

• Is (A) alone sufficient, deferring (B) indefinitely? What use cases actually require serving-time audio→words?
• What is the minimal lawful phonetic feature set (formants? MFCC-like bands? — noting MFCC must stay deterministic and inspectable, not a learned frontend)?
• How are acoustic↔lexeme associations represented in a pack, and how are they taught/reviewed/versioned/checksummed?
• What is the honest accuracy ceiling of (B), and how is it measured (a sealed audio eval lane analogous to the GSM8K sealed test)?
• Speaker/accent/noise robustness without learned models — scope or explicitly out-of-scope for v1?
• Does (B) compose with the binding-graph / comprehension reader the same way the text path does, so downstream is unchanged?

6. Decision

Deferred. This stub records the fork and the constraints. No path is selected yet. The serving path remains Whisper-free and audio remains paralinguistic until a full ADR selects and specifies (A) and/or (B).

7. Consequences

• Until this is taken up, audio comprehension is prosody/turn/affect only at serving time; lexical content for audio is unavailable in production (text modality carries words).
• Recording the fork now prevents the silent failure mode: someone wiring a teacher into the serving path because "audio needs words" without confronting the doctrinal cost. That decision must go through this ADR's successor.

8. References

• ADR-0181 — Audio compiler; teacher/shadow policy; embeddings rejected as substrate.
• ADR-0180 — Delta-CRDT substrate; trace/merge determinism the lexeme path must preserve.
• ADR-0164 / ADR-0165 — reviewed growth of lexicon entries and lexeme primitives (the corridor a taught audio vocabulary would reuse).
• docs/audio_pipeline_overview.md §9 — teacher = scaffolding; serving path stays Whisper-free; paths (A)/(B).
• docs/model_dependency_size_tally.md — the 0-parameter tally this path must not move.