AAA-somatic-emd-pipeline

AAA · Somatic EMD Pipeline

Acoustic_intent without words. Same constitutional floor. Killed by the same barge-in. Voice-cadence proxy now; WELL biometrics when sensors live; F13 for state-change claims.

DITEMPA BUKAN DIBERI.

The Three-Phase Reflex Arc (music)

        ┌─────────────────────────────────────────┐
        │  PHASE 1 — DECODE (Sense)                │
        │   Voice waveform → somatic_proxy         │
        │   Source: WPM, pitch_mean,               │
        │   hesitation_ms, RMS variance            │
        │   Label: OBS (cadence) + INT (state)     │
        │   NOT WELL until Layer 2                 │
        └────────────────┬────────────────────────┘
                         │
                         ▼
        ┌─────────────────────────────────────────┐
        │  PHASE 2 — METABOLIZE (Route)            │
        │   somatic_proxy → music_intent packet   │
        │   {role, BPM_target, density,            │
        │    max_seconds, kill_on_barge_in}        │
        │   NO melody choice. NO lyric choice.     │
        │   Just acoustic parameters.              │
        └────────────────┬────────────────────────┘
                         │
                         ▼
        ┌─────────────────────────────────────────┐
        │  PHASE 3 — ENCODE (Render)               │
        │   music_intent → MiniMax T2A Music       │
        │   Parametric: BPM, key, texture, density │
        │   Lane 6 in catalog (Layer 0 = lane 1-5) │
        │   Hard i-ARIF DENY for vocal melodic     │
        └─────────────────────────────────────────┘

This arc runs in parallel with the voice EMD pipeline (AAA-audio-emd-pipeline). The two arcs share acoustic_intent envelope but emit to different encoder lanes. The packet below shows both.

The Unified Packet (voice + music)

{
    # Existing voice fields (from AAA-audio-emd-pipeline)
    "text": "Hang biar betul?",
    "epistemic_label": "INT",
    "acoustic_intent": {
        "emotion": "calm-confident",
        "intensity": 0.6,
        "pace_wpm": 165,
        "pitch_delta_hz": -8,
        "breath_tokens": ["[breath]", "[uv_break]"],
    },

    # NEW: Somatic proxy (Layer 1 source — voice cadence)
    # GUARDRAIL: epistemic_label=INT mandatory (never OBS for derived state)
    "somatic_proxy": {
        "source": "voice-cadence",       # explicit (not WELL until Layer 2)
        "epistemic_label": "INT",        # GUARDRAIL: derived, never OBS
        "wpm": 172,                       # OBS — measured
        "pitch_mean_hz": 192,             # OBS — measured
        "hesitation_ms": 45,              # OBS — measured (low = rushed)
        "rms_variance": 0.18,             # OBS — measured (high = intense)
        "derived_state": "elevated",      # INT — F7 cap 0.75 (elevated is least certain)
        "derived_confidence": 0.71,       # GUARDRAIL: F7-capped per state class
        "derivation_basis": "wpm>180 + hesitation<200ms + pitch>200Hz",  # GUARDRAIL: must cite measurements
    },

    # NEW: Music intent (this skill's contribution)
    # GUARDRAIL: semantic_route_audit mandatory on every emission
    "music_intent": {
        "role": "bed",                    # silence | bed | warn | none
        "bpm_target": 70,                 # derived from WPM / ~2.3
        "key": "auto",                    # engine chooses within musical bed
        "density": "low",                 # very_low | low | medium | high
        "max_seconds": 8,                 # hard cap (F1)
        "kill_on_barge_in": true,         # mandatory on voice-mode path
        "no_vocal": true,                 # no lyric, no vocal melodic
        "no_i_arif": true,                # hard DENY for i-ARIF voice_id
        "confidence": 0.71,               # GUARDRAIL: F7 cap, matches proxy confidence
        "reason": "wpm=172 derived elevated → down-regulate via breath-rate match",  # GUARDRAIL: derivation chain
        "reversible": true,               # GUARDRAIL: always reversible
        "semantic_route_audit": "wpm→bpm direct derivation, no emotion inference",  # GUARDRAIL: audit trail
    },

    # Falsification gate (Layer 2 — placeholder for now)
    "falsification": {
        "voice_round_trip": null,         # STT round-trip on TTS (existing)
        "biometric_round_trip": null,     # WELL feedback (Layer 2 only)
        "rejected": False,
    },

    "policy": {
        "verdict": "ALLOW",               # ALLOW | HOLD | DENY
        "f13_token": null,                # required for Layer 3 / sustained
        "reason": "voice-cadence elevated → bed 70 BPM low density",
    },

    "receipt_id": "abc123..."             # F11 audit
}

Phase 1 — DECODE (somatic_proxy from voice cadence)

Layer 1 source (live NOW)

# AAA-asr-glm-ingest / AAA-audio-emd-pipeline produces these fields.
# GUARDRAIL (2026-09-04): epistemic_label=INT, derivation_basis, confidence≤0.75
somatic_proxy = {
    "source": "voice-cadence",
    "epistemic_label": "INT",              # GUARDRAIL: never OBS for derived state
    "wpm": compute_wpm(last_n_utterances),         # OBS
    "pitch_mean_hz": compute_pitch_mean(waveform),  # OBS
    "hesitation_ms": compute_pause_distribution(),  # OBS
    "rms_variance": compute_rms_variance(waveform),  # OBS
    "derived_state": derive_state_int(...),          # INT, F7 ≤ 0.75
    "derived_confidence": F7_cap(0.75),              # GUARDRAIL: lowered from 0.90
    "derivation_basis": f"wpm={wpm},pitch={pitch},hes={hes}",  # GUARDRAIL: cite measurements
}

Labels are non-negotiable:

• wpm, pitch_mean_hz, hesitation_ms, rms_variance → [OBS] (raw measurement).
• derived_state, derived_confidence → [INT] (interpretation, F7 capped).
• The whole object is labeled somatic_proxy — never hang_state, user_stress, arif_mood.

Honest derivation rules (F7 cap — updated 2026-09-04):

Two listeners hear two intents. Any agent that reports derived_state with confidence > cap has overclaimed. Return VOID.

GUARDRAIL (2026-09-04): Every derived_state MUST carry derivation_basis citing which OBS measurements support the INT. No basis = VOID.

Layer 2 source (HOLD until WELL sensors live)

# PLACEHOLDER — NOT YET WIRED.
somatic_proxy = {
    "source": "well-biometric",
    "hrv_rmssd": ...,                   # OBS — when WELL sensor live
    "heart_rate_bpm": ...,              # OBS
    "skin_conductance": ...,            # OBS
    "sleep_debt_hours": ...,            # OBS
    "derived_state": derive_state(...),
}

Activation rule: Layer 2 source can only be enabled after:

1. WELL biometric sensor hardware is live (sensor-debt currently 111 days).
2. A-FORGE Paradox Engine wired to arifOS kernel.
3. Falsification gate has run on at least 30 days of cross-correlated data.
4. human_approval_token to close the loop.

Until all four: Layer 1 voice-cadence is the ONLY source. No WELL data in the packet.

Phase 2 — METABOLIZE (harmonic router)

The metabolizer's job is NOT to compose. Its job is to map somatic_proxy → music_intent parameters. No melody choice, no lyric, no arrangement. Pure acoustic parameter routing.

Routing Table (Layer 1)

def metabolize_music_intent(somatic_proxy: dict) -> dict:
    """Layer 1: voice-cadence → music_intent. Pure routing, no composition."""

    wpm = somatic_proxy["wpm"]
    state = somatic_proxy["derived_state"]

    # F1 fail-closed
    if somatic_proxy.get("source") != "voice-cadence":
        return _silence_intent(reason="telemetry source not allowed for Layer 1")

    # Honest BPM derivation (not sacred Hz tables)
    if state == "elevated" or wpm > 180:
        return {
            "role": "bed",
            "bpm_target": 70,         # down-regulate via breath-rate match
            "key": "auto",
            "density": "low",
            "max_seconds": 8,
            "kill_on_barge_in": True,
            "no_vocal": True,
            "no_i_arif": True,
            "rationale_int": "elevated WPM suggests down-regulation via breath-rate match",
        }

    if state == "focused":
        return {
            "role": "bed",
            "bpm_target": int(wpm / 2.3),   # match cadence, not sacred tables
            "key": "auto",
            "density": "low",
            "max_seconds": 30,              # sustained focus, F11 audit
            "kill_on_barge_in": True,
            "no_vocal": True,
            "no_i_arif": True,
            "rationale_int": "focused WPM suggests cadence-matching bed",
        }

    # Default: silence (don't emit arbitrary music)
    return {
        "role": "silence",                   # F1: don't fabricate state
        "bpm_target": None,
        "max_seconds": 0,
        "rationale_int": "no state signal → silence (fail-closed)",
    }


def metabolize_f1_warning(somatic_proxy: dict, risk_class: str) -> dict:
    """F1 / high-risk: dissonance as pre-lexical alert. NO medical claim."""

    if risk_class not in {"F1_AMANAH", "F13_SOVEREIGN", "high-risk"}:
        return _silence_intent(reason="risk_class not authorized for warn")

    return {
        "role": "warn",
        "bpm_target": None,                  # dissonance is rhythm-agnostic
        "key": "auto",
        "density": "high",                   # spectral roughness ↑
        "max_seconds": 3,                    # pre-lexical only
        "kill_on_barge_in": True,
        "no_vocal": True,
        "no_i_arif": True,
        "rationale_spec": "F1 trigger → 3s spectral roughness, no lyric",
    }

Hard rules in the metabolizer

1. No lyric output. Music with lyric in the voice-mode flow competes with the human's speech processing. Default is no_vocal=True and stays so unless explicit F13 override.
2. No melody choice. The agent picks BPM, density, duration. The engine picks key, timbre, arrangement. Separation of concerns.
3. No i-ARIF clone. If voice_id resolves to i-ARIF, the music encoder MUST DENY. i-ARIF singing is forbidden per F13 + DENY.
4. No silence fabrication. When state signal is missing, return role: silence, NOT a default bed. F1 fail-closed.

Phase 3 — ENCODE (parametric music synthesis)

The encoder routes music_intent to a parametrically-controllable music engine. The default lane is MiniMax T2A Music (music-2.6 or music-3.0).

Engine routing

Example (MiniMax T2A Music call)

import requests

def encode_music_intent(intent: dict, api_key: str) -> bytes:
    """Phase 3: emit authorized waveform."""

    # F1 fail-closed
    if intent["role"] == "silence":
        return b""

    # F13 / i-ARIF DENY
    if not intent.get("no_i_arif", False):
        return VOID  # policy gate

    # F1 frequency / dB bounds
    if intent["max_seconds"] > 30:
        return VOID  # F13 territory

    payload = {
        "model": "music-2.6",                # or music-3.0
        "prompt": _build_prompt(intent),     # see below
        "bpm": intent["bpm_target"],
        "duration": intent["max_seconds"],
        "format": "mp3",
        "sample_rate": 24000,
    }

    # CRITICAL: separate endpoint from voice clone
    # POST /v1/music_generation (not /v1/voice_clone)
    r = requests.post(
        "https://api.minimax.io/v1/music_generation",
        headers={"Authorization": f"Bearer {api_key}"},
        json=payload,
        timeout=30,
    )
    r.raise_for_status()
    return r.content


def _build_prompt(intent: dict) -> str:
    """Translate intent → text prompt. NO lyric, NO melody choice."""
    if intent["role"] == "bed":
        density = {
            "very_low": "extremely sparse, single sustained tone, no rhythm",
            "low": "low entropy, minimal rhythm, ambient texture, no vocal",
            "medium": "moderate texture, subtle pulse, ambient, no vocal",
        }[intent["density"]]
        return f"Ambient bed, {intent['bpm_target']} BPM, {density}, no vocal"
    if intent["role"] == "warn":
        return f"Dissonant alert tone, spectral roughness, 3 seconds, no vocal"
    return "silence"

Hard rules in the encoder

1. Endpoint separation — /v1/music_generation is the music lane. /v1/voice_clone is the i-ARIF lane. They MUST NOT cross. Music payload to clone endpoint = identity leak (F1).
2. Format — mp3 for streaming, pcm16 only for true realtime pipe. wav for archival.
3. Timeout — 30 s max per emission. If engine hangs, kill and return VOID.
4. Reception — receive response, write to F11 audit log with receipt_id, then route to playback.

Kill-on-Barge-in (the same gate)

Music playback is killed by the same barge-in signal that kills voice playback. Same voice.barge_in = true config key. Same RMS threshold. Same grace period.

# EMD Phase 3 must register a kill listener on the barge-in bus.
def on_barge_in():
    halt_music_playback()        # cut current emission
    cancel_pending_emission()    # cancel queued emissions
    return_to_silence()          # F1 fail-closed

When kill_on_barge_in=True (default for voice-mode flow), the encoder registers this listener before emission. The music lane inherits the same physics-of-conversation floor as the voice lane.

Policy Gates (in voice_filters.py extension)

Layer 1 ignite adds these checks to /root/.hermes/voice_filters.py:

def check_somatic_policy(intent: dict, voice_id: str) -> PolicyVerdict:
    """F1 + F13 + i-ARIF DENY gates for somatic music."""

    # F1: hard fail-closed
    if intent.get("max_seconds", 0) > 30:
        return PolicyVerdict(allowed=False, reason="F1: duration > 30s without F13")

    # F1: no medical claim
    if any(kw in str(intent).lower() for kw in
           ["heal", "cure", "therapy", "therapeutic", "diagnose"]):
        return PolicyVerdict(allowed=False, reason="F9: medical claim detected")

    # F13: i-ARIF singing DENY
    if voice_id.startswith("i-ARIF") and intent.get("no_vocal") is False:
        return PolicyVerdict(allowed=False, reason="F13: i-ARIF vocal melodic DENY")

    # F1: group lane stays Sado-flat — no music overlay
    if voice_id == "ttv-voice-2026081808404926-BdoQh6ec":
        return PolicyVerdict(allowed=False, reason="F1: Sado-locked group lane, no music overlay")

    # F11: audit + receipt
    receipt = _receipt("somatic.allow", {"intent_role": intent.get("role"), "voice_id": voice_id})
    return PolicyVerdict(allowed=True, reason="Layer 1 voice-cadence", receipt_id=receipt)

Layer 1 / Layer 2 / Layer 3 Activation Matrix

Integration with Voice EMD

The music EMD arc runs alongside the voice EMD arc, sharing acoustic_intent envelope but emitting to a different lane:

Voice Mode conversation turn
   ↓
Phase 1 (shared) — voice cadence → somatic_proxy
   ↓
   ├── Voice EMD continues (Phase 2 → Phase 3 → TTS)
   └── Music EMD continues (Phase 2 → music_intent → Phase 3 → T2A Music)
   ↓
   Both outputs gated by:
     - barge-in (kill both)
     - voice_id policy (i-ARIF DENY for music)
     - group lane policy (Sado lock, no music overlay)
     - F11 audit (separate receipt per emission)

When to Load This Skill

• Any agent adds a music lane to a Voice Mode conversation.
• Any agent emits ambient frequency during dialogue.
• Any agent routes music_intent to MiniMax T2A Music.
• Any agent considers WELL biometric wiring (Layer 2).
• Any agent debugs a "music didn't stop" / "music overlapped Sado" issue.
• Any agent audits an emission that claims state-change.

Integration Points

• Doctrine (this family): /root/AAA/skills/AAA-somatic-music-doctrine/SKILL.md
• Engines (this family): /root/AAA/skills/AAA-somatic-engine-catalog/SKILL.md
• Voice EMD: /root/AAA/skills/AAA-audio-emd-pipeline/SKILL.md
• ASR / voice cadence: /root/AAA/skills/AAA-asr-glm-ingest/SKILL.md
• i-ARIF DENY: /root/AAA/skills/AAA-voice-cloning-mimo-minimax/SKILL.md
• Composer lens (untouched): /root/HERMES/skills/media/music-intelligence/SKILL.md
• Enforcement module: /root/.hermes/voice_filters.py (adds somatic lane in Layer 1)
• WELL bridge (Layer 2): mcp__well__well_assess_homeostasis
• A-FORGE Paradox Engine (Layer 2): /root/A-FORGE/paradox-engine/

Related Skills

• AAA-somatic-music-doctrine — Three-layer maturity + corrected thesis
• AAA-somatic-engine-catalog — Parametric engine registry
• AAA-audio-qualia-doctrine — F10/F9 for voice (sister doctrine)
• AAA-audio-emd-pipeline — Voice EMD reflex arc (sister pipeline)
• AGI-audio-quantum-cognition — Audio physics + floors (parent)
• music-intelligence — Composer lens (Layer 0, untouched)
• AAA-voice-cloning-mimo-minimax — i-ARIF DENY surface
• well_assess_homeostasis — Layer 2 substrate (HOLD)

Pipeline forged 2026-08-18. F2 evidence: derivation from ARIF's corrected framing (voice-cadence proxy, music_intent packet field, kill-on-barge-in shared gate, i-ARIF singing DENY, honest BPM derivation) + AAA-audio-emd-pipeline v1.0.0 + AGI-audio-quantum-cognition v1.0.0. F1 floor absolute: fail-closed to silence when telemetry missing. F9 floor absolute: agent is instrument, not shaman.