ADR 0007: U3 AcousticPlane — Symbolic & Spectral-First Audio

Status

Accepted (2026-06-17)

Context

QualiaDB’s phenomenal viewport (U2) projects the shared [α, μ, σ] truth layer from Tensor10D SOA into linear RGB via CIE Gaussian CMFs. Users expect the same semantic manifold to be heard as well as seen — without introducing a second truth stack or routing neural PCM from the LLM inference universe (U0).

Competing approaches considered:

1. LLM waveform generation — U0 emits PCM buffers. Rejected: blows 42 MB Sentinel budget, adds non-deterministic latency, violates zero-heap hot paths.
2. <audio> tag + MP3 URLs — Simple but lossy; psychoacoustic discard destroys sovereignty over dynamic range (α).
3. Web Audio API from JS — Reintroduces V8 spectral math the migration plan retires.
4. Symbolic tokens + spectral sheets + parametric DSP — Matches visual cold/hot split (bake → resident → render).

Browser constraints add a transport decision: AudioWorklet runs off the main thread; uniform updates must be either postMessage or SharedArrayBuffer (requires COOP/COEP).

Decision

Introduce ComputeUniverse::AcousticPlane (U3) as a read-only alias of the U1 Tensor10D ledger partition with its own hot-path consumer:

Locked invariants

1. U0 never emits PCM — only Sonic Tokens and graph structure.
2. σ parity — visual λ (400–700 nm) and auditory Hz (1760–110) derive from the same fract(σ) mapping (portal_acoustic.rs).
3. U3 muted in Reserve — acoustic synthesis yields to LLM KV retention under VRAM pressure.
4. Analytic HRTF in hot path — KEMAR tables, if adopted, are cold-path assets only.
5. Sonic Token layout — 8-byte Pod; tensor index at bits 32–47; SONIC_MAGIC 0x53 in flags.

Consequences

Positive

• Phenomenal multi-modal fidelity: orbit camera → pan moves; select node → σ drives timbre in both modalities.
• Zero-copy uniform path on COI-enabled Pages deploys.
• CI oracles (phenomenal_sigma_visual_audio_parity, phenomenal_hrtf, 22 audio:: tests) prevent layout drift.
• Same resident tensor buffer feeds U2 draw calls and U3 binaural position — no duplicate scene graph.

Negative

• COOP/COEP requires service worker + reload on GitHub Pages; users without isolation still work but pay MessagePort copy cost.
• Full CQT mmap ingest and KEMAR HRTF deferred — analytic model is less spatially accurate than measured HRTF.
• AUDIO_PROJECT_STATUS.md and webizen audio_contract.rs salvage docs required migration to qualiaDB ownership.

Neutral

• WASM portal size grew (~436 KB raw); still within slim portal budget after gzip.
• Desktop host (webizen-browser) does not yet mirror U3 — Pages path is golden master until PR-C10.

References

• docs/manuals/qualia-wasm-portal.md
• docs/manuals/standards/q42-acoustic-plane-draft.md
• docs/plans/wasm-viewport-migration-plan.md — Track B5, P-F1/P-F2
• ADR 0001 — 48-byte Quin alignment (Sonic Tokens are separate 8-byte events, not Quins)