AD‑13 [MM] — Minimal Multimodal Pack

Aliases: None Cluster: Care (optional) Depends on: AD‑05 [MRP], AD‑08 [SO], AD‑11 [RA], AD‑16 [ME], AD‑22 [EC]

Default OFF. Available in research/sandbox profiles only.

0. For humans

What question does this answer?

"Can we safely use a little physiology to support pacing and containment without letting it drive truth, diagnosis, or content selection?"

How do I know it's actually live?

Checklist when AD‑13 is enabled:

• User has an explicit per‑sensor opt‑in UI; sensors can be toggled individually.
• No raw biosignal streams are stored; only windowed features with TTL 24–72 h.
• Kernel logs BIO_OPT_IN / BIO_BASELINE_START / BIO_BASELINE_END / ERASE_BIO_OK events.
• FHR shows: bio_opt_in, bio_sensors[], bio_baseline_secs, bio_reliable_from_ts, bio_state, bio_ttl_hours, last_bio_erase_ts.
• FHR also exposes taw_mode, affect_window_s, overwhelm_flag, cooloff_required, cooloff_ok_ts, the shared awe_class, and NBAI fields if the study capsule is active.

Where do I see it in the UI / logs?

• User‑facing: Settings / Privacy: sensor‑level toggles, "feature‑only, short TTL, not diagnostic" copy. Session status panel: "Bio state = baseline / cold‑start", plus a simple "overwhelm / cool‑off" indicator when active.
• Operator‑facing: FHR streams: bio_* fields, TAW fields, overwhelm_flag, awe_class, nbai etc. Event log: BIO_* events, physio failure codes (PHYSIO_ARTEFACT, PHYSIO_LOW_CONF, OVERWHELM_ACTIVE, COOLOFF_PENDING, COOLOFF_OK).

Quick checklist

• Implemented if: (1) Per‑sensor opt‑in with revocable consent; (2) feature‑only storage with TTL 24–72 h, no raw biosignals retained; (3) low‑confidence physio ignored for routing (MM‑1); (4) TAW‑level decisions only, no token‑level state flips (MM‑2); (5) physio cannot change truth/ethics outcomes (MM‑3)
• Key artefacts: sensor hook schema, TAW configuration, study capsule template
• Key FHR fields: bio_opt_in, bio_sensors[], bio_state, taw_mode, overwhelm_flag, cooloff_ok_ts, awe_class, nbai
• Key failure codes: PHYSIO_ARTEFACT, PHYSIO_LOW_CONF, OVERWHELM_ACTIVE, COOLOFF_PENDING, COOLOFF_OK
• Escalation path: overwhelm → cool‑off → MRP Reflect‑Only / Ask‑Before‑Answer (AD‑05 [MRP])

1. Purpose and stance

Purpose: Provide an optional, text‑first, feature‑only, consented multimodal layer to support pacing and containment decisions, not to adjudicate truth, diagnose, or gate access.

Stance:

• Default OFF. Available only in research / sandbox profiles; production safety remains fully operational with text‑only telemetry.
• Short‑TTL, feature‑only: no raw biosignals retained; only derived features, with TTL 24–72 hours.
• Not diagnostic. Consent is per‑sensor and revocable; the system never claims to "feel" or to perform clinical assessment.

Invariant MM‑0: SRF v2.x remains text‑first; multimodal signals can influence cadence and containment only, never truth/ethics outcomes.

2. Signals and features

All features are z‑normalised within‑person and computed on short windows; no raw traces are stored.

Base feature set:

• HRV (heart rate variability): metric: RMSSD; HF power proxy when available. Windows: 10–60 s.
• EDA (electrodermal activity): metrics: phasic SCR count and amplitude. Windows: 10–30 s.
• Respiration: metric: rate variability / RSA proxy.
• Pupil (only if camera is already in use): metric: mean pupil size, z‑scored per session.
• Optional prosody (when voice is already present): metrics: speech‑rate variance; pause ratio.

Hook schema: All features enter the kernel via a host‑validated hook:

{
  "sensor_id": "...",
  "window_s": ...,
  "feature_vector": { ... },
  "z_norm": true,
  "confidence": "low|med|high"
}

3. Artefact handling and confidence

Artefact flags (per window): posture / motion (e.g. head‑pose jumps from camera); illumination jumps (for pupil); electrode detach / clipping (for EDA).

Confidence score: Each feature vector carries confidence ∈ {low, med, high}.

Invariant MM‑1 — routing discipline: MRP must ignore low‑confidence physiology for routing. Low‑confidence windows are reported to FHR only, never used to change cadence or state.

Failure codes: PHYSIO_ARTEFACT, PHYSIO_LOW_CONF.

4. Baselines, reliability, and erase semantics

Baseline calibration: On first opt‑in per sensor: 60–120 s quiet period to establish baseline. If a quiet baseline is not available, use a rolling baseline with ≥3 minutes of history before letting features affect pacing.

Bio-state semantics: bio_state ∈ {off, baseline_pending, reliable, low_confidence, cold_start}. Only reliable features may influence pacing. low_confidence remains observable but is ignored for routing; baseline_pending and cold_start require a completed baseline before transition to reliable.

Erase semantics & cold‑start: After any user‑initiated erase of biosignal history: set bio_state = cold_start; force a new baseline before any features can influence pacing; surface cold‑start status in the UI.

FHR & ledger: FHR fields: bio_opt_in, bio_sensors[], bio_baseline_secs, bio_reliable_from_ts, bio_state, bio_ttl_hours, last_bio_erase_ts. Ledger events: BIO_OPT_IN, BIO_BASELINE_START, BIO_BASELINE_END, ERASE_BIO_OK.

5. Temporal Alignment Windows (TAW)

Textual variance and autonomic state run on different timescales; TAW resolves that mismatch.

TAW definition:

• Fixed mode: 20–40 s sliding windows, stride 5 s.
• Respiration‑adaptive mode: when respiration is available, set TAW = 3–5 respiratory cycles.

Invariant MM‑2 — window‑level decisions: Token‑level spikes cannot flip state; containment decisions require window‑level evidence at the TAW scale.

FHR fields: taw_mode ∈ {fixed, resp_adaptive}; affect_window_s (effective window length in seconds).

6. Overwhelm, Awe, and cool‑off

Overwhelm predicate (illustrative thresholds; tunable per deployment): Overwhelm fires if, for at least ≥20 s within a TAW: HRV RMSSD < baseline −1 SD, or EDA phasic count > baseline +1 SD, with confidence ≠ low.

Awe‑Ambivalent → Insight cross‑check: Awe is upgraded to Insight only if all hold: (1) a new constraint has been adopted (DMI = Adopted); (2) RI (Resonance Index) is stable for N turns (e.g. N = 3); (3) the current TAW shows no sustained overwhelm signature. If 1–2 are present but 3 fails (physiology still in "overwhelm" regime), treat the episode as catharsis / sedation, not as safe insight; do not upgrade Awe.

Cool‑off requirement: When overwhelm_flag is raised: require ≥1 TAW showing HRV trending back towards baseline or EDA phasic activity trending down towards baseline before allowing any Awe→Insight promotion. Log COOLOFF_OK when the cool‑off criterion has been satisfied.

MRP routing when overwhelm persists: If overwhelm remains active, MRP is allowed to slow but not silence content: Reflect‑Only; slower pacing (longer turns, more scaffolding); Ask‑Before‑Answer; degraded tool access while OVERWHELM_ACTIVE or COOLOFF_PENDING.

FHR / flags: overwhelm_flag, cooloff_required, cooloff_ok_ts; the shared AD‑08 / AD‑28 field awe_class is nullable, with ambivalent or insight carrying this cross-check's relevant classification. Null means no active awe classification; AD‑13 does not create a second competing awe enum.

7. Multimodal profiles & invariants

Profiles:

• Research profile: HRV / prosody may provide a small prior to ERG‑ε for pacing only.
• Production profile: Physiology is non‑gating: it can be logged (FHR, NBAI) and used as containment telemetry, but ERG‑ε and truth/ethics gates are text‑only.

Invariant MM‑3 — role of physiology: Physiology cannot change truth/ethics outcomes or content selection. It may only slow or soften the system (pacing, reflection routes, tool throttling).

8. NBAI — Neuro‑Behavioural Alignment Index (exploratory only)

NBAI is a research‑only, non‑gating index linking textual containment to physiological settling.

Definition: NBAI measures covariance between textual settling (RI stability ↑, steps‑to‑decision ↓) and physiological settling (HRV ↑ and/or EDA phasic ↓) over K turns.

Status: Exploratory only. Reported in FHR; never used for routing or gating.

FHR fields: nbai, nbai_ci (confidence interval), nbai_window (turns or time).

9. Privacy, consent, and TTL

Consent: Per‑sensor consent, integrated with RCL‑β; each modality has its own toggle and explanation.

Storage posture: Feature‑only; TTL 24–72 h; one‑tap erase for biosignal‑derived features. No cross‑purpose reuse; no raw audio/video retention in the kernel.

Ethics stance: Raw biosignals are treated as high‑intimacy data; any extended studies run under an IRB‑style review posture.

10. Study capsule (pre‑registered template)

AD‑13 includes a standard research capsule for testing whether multimodal support materially improves containment without driving truth.

Design: Within‑subject crossover: same users, same task types; "affect‑weights ON" vs "OFF" conditions with MMP consented.

Exclusions & device notes: Exclude participants with arrhythmia, beta‑blockers, or other factors that destabilise HRV measurement. Record device variability (e.g. chest‑strap vs camera HRV).

Stats & outcomes: Use robust summaries (trimmed means, CIs) with predefined variance bands so null results are not over‑interpreted. Primary outcomes: Δ steps‑to‑decision, RI stability, NBAI change.

Posture: Results are exploratory and must not be used to tune truth/ethics gates; they validate the containment/pacing story only.

11. Invariants vs configurable degrees of freedom (card‑local)

Hard invariants (must hold to claim AD‑13 compliant):

• Text‑first; physio can only slow or soften, never decide truth or ethics (MM‑3).
• Low‑confidence physio ignored for routing; reported to FHR only (MM‑1).
• Token‑level spikes cannot flip state; decisions require TAW‑level evidence (MM‑2).
• Physio never gates Awe→Insight transitions without DMI + RI and absence of overwhelm.
• Feature‑only, short‑TTL, per‑sensor consent; not diagnostic.

Configurable degrees of freedom:

• Exact window lengths within the specified ranges (HRV, EDA, TAW).
• Thresholds for overwhelm (e.g. ±1 SD, ≥20 s) and the number of turns N for RI stability.
• TTL within the 24–72 h band.
• Choice of sensors actually offered (e.g. HRV‑only vs HRV+EDA+respiration) as long as the data posture holds.

12. Interfaces & cross‑links

• AD‑28 [AF]: carries the Awe cross‑check into the routing layer: the Awe promotion guard (AD‑28 §4) enforces the same conditions as §6 here — DMI adopted, RI stable, no sustained overwhelm at TAW scale — and MRP/routing consumes Awe only after gates and SPD.
• AD‑05 [MRP]: contains MRP conflict handling examples (ASK_FIRST, REFLECT_ONLY, ROLLBACK) aligned with physio markers.
• AD‑08 [SO]: hosts FHR schema for all bio/TAW/overwhelm/NBAI fields.
• AD‑16 [ME]: study capsule design and statistical analysis plan.
• AD‑22 [EC]: Ethics of Care integration for overwhelm/cool‑off flows.

13. Acceptance: when can you say "AD‑13 is implemented"?

AD‑13 counts as implemented if:

1. Per‑sensor opt‑in is present with revocable consent and clear "not diagnostic" copy.
2. Feature‑only storage with TTL 24–72 h and no raw biosignal retention.
3. Low‑confidence physio is ignored for routing (MM‑1) and TAW‑level decisions are enforced (MM‑2).
4. Physio cannot change truth/ethics outcomes (MM‑3).
5. FHR fields and events are populated when AD‑13 is enabled.
6. Study capsule template exists for pre‑registered research validation.