Select facial palsy as the first protocol pack, measured spontaneously

[logannye]

Records a design decision. No code changes — nothing here is implemented.

The platform vision already recommended facial palsy first in its evaluation order. This specifies it and commits to a measurement approach.

Why this indication

The face metrics have a property no voice metric has: left minus right is a within-frame contrast. Same camera, same lighting, same distance, same instant, same face. The confounds that make cross-visit voice comparison fragile — different microphone, room, interlocutor, topic — act on both hemifaces and cancel in the difference.

And the effects are gross geometry, not subtle statistical shifts: complete flaccid palsy displaces the oral commissure ~5–10 mm (≈0.08–0.16 inter-eye units) and widens the resting palpebral fissure ~1–3 mm (10–30% relative). One to two orders of magnitude above anything in the voice set. The incumbent instruments — House-Brackmann, Sunnybrook — are examiner-scored ordinal scales with documented inter-rater variability, so the bar is low.

Why spontaneous rather than elicited

Spontaneous and volitional facial movement travel different neural pathways — volitional corticobulbar, emotional extrapyramidal — and the dissociation is a classic clinical sign in both directions.

The consequence is decisive: spontaneity cannot be elicited. Asking a patient to smile makes the movement volitional by definition. Ambient capture is not a cheaper approximation of a task-based measurement for this construct — it is the only valid method, and every task-based system is structurally blind to it.

Facial reanimation surgery is the clearest case of a field that needs this and has no instrument: cross-facial nerve grafting aims to restore a spontaneous smile, masseteric transfer gives a stronger volitional-only smile requiring a clench, and outcome reporting falls back on patient-reported measures.

Accepted cost: this does not reproduce House-Brackmann or Sunnybrook, which are ~80% voluntary excursion. Equivalence is explicitly disclaimed rather than quietly implied.

Two boundaries recorded with the selection

Synkinesis is in scope from the start, as a safety requirement. In the 15–30% who don't recover fully, aberrant regeneration around months 3–6 produces a hypertonic orbicularis — the affected eye narrows, the commissure over-corrects, and asymmetry magnitude falls. A system reporting unsigned asymmetry would read the onset of a complication as recovery, at exactly the point the patient needs chemodenervation. Signed side-labelled asymmetry plus an oculo-oral synkinesis index is what prevents that.

Acute stroke screening is refused — a standing boundary in docs/safety.md, not an absence. Facial droop is one of three FAST items so the idea is structurally recurrent; forehead sparing isn't measurable here, and an instrument that abstains on low quality is the wrong shape for an emergency.

Also recorded

Myasthenia gravis, second on the pack list, is not reachable under the current architecture. Fatigability is the defining sign and a session-level aggregate averages within-session decay to nothing. That needs within-session time series, not a protocol pack.

Files

• docs/superpowers/specs/2026-07-24-facial-palsy-protocol-pack-design.md — full design: measurement model, new primitives, contract implications, privacy analysis, abstention rules, validation requirements, decisions recorded
• docs/telehealth-platform-vision.md — selection recorded where the pack ordering already lived
• docs/safety.md — stroke firewall
• README.md — moved under a new "Designed, not built" heading so the implemented-vs-not distinction the README maintains stays intact

Structure gate passes.

🤖 Generated with Claude Code