# AI Translator for Healthcare — UI Flow Design **Product:** Clinically validated, HIPAA-ready real-time translator for bedside, intake, and telehealth. **Design goal:** Clear, fast, trustable communication that a nurse can pick up in 30 seconds and a patient can use without being able to read. **Companion doc:** [`RESEARCH.md`](RESEARCH.md) — competitor + UX research that informs every decision here. --- ## 1. Hero Flow Choice **Primary flow designed in detail:** Bedside conversation. **Supporting flows:** Triage / intake (kiosk), Review & export. | Flow | Why included | |---|---| | **Bedside** (hero) | Highest-volume, biggest pain (waiting for human interpreter mid-rounds). Best showcase of Clinical Clarity Filter + face-to-face split + human escalation. | | **Intake (kiosk)** | Demonstrates speed + auto language detection + structured data handoff to chart — the "delay reduction" benefit. | | **Review & export** | Demonstrates the audit trail + compliance signal — the "HIPAA-ready" benefit. | Three flows together cover the brief's three benefits: **fewer errors** (bedside + Clarity Filter), **less delay** (intake + AI-first), **compliance** (review/export). --- ## 2. Target Platforms & Form Factors Designing for healthcare means picking platforms hospital IT will approve, clinicians already own, and HIPAA reviewers will sign a BAA on. Decisions below; the harder calls are deferred to [§8 Open Questions](#8-open-questions-for-the-stakeholder). ### Launch matrix | Platform | Role | Form factor | Why | |---|---|---|---| | **iOS native** (Swift / SwiftUI) | **Primary** | iPhone (clinician pocket) + iPad (bedside, kiosk) | ~70% of US hospital-issued clinician phones are iOS; on-device Speech framework + Core ML for sub-300ms latency; FaceID for biometric re-entry; tight Bluetooth/AirPods integration; App Store rigor aligns with healthcare procurement | | **Responsive web** (Next.js + Tailwind, SMART-on-FHIR embeddable) | **Co-launch** | Workstation, EHR embed | Launch-from-chart inside Epic Hyperspace / Cerner PowerChart without leaving the EHR; runs on any clinician workstation; no app-store delay for hospital IT | | **Android** | Post-MVP (Q3+) | Phone / tablet | ~30% of US healthcare devices but harder HIPAA story across fragmented OEMs. Defer until iOS + web prove the wedge. | | **Apple Watch** | Explored, deferred | Wrist | Glanceable transcript + tap-to-pause attractive, but watch radio bandwidth isn't reliable enough for live audio yet. | ### Form factor per flow - **Screen 4 — bedside (hero):** iPhone or iPad held flat between clinician and patient. The inverted top half assumes physical device orientation, *not* a software split view — design tokens differ between iPhone (portrait, 6.1") and iPad (landscape, 11"). - **Intake kiosk (bonus):** iPad mounted on a registration stand. Patient is standing; tap targets bumped to **≥ 56pt**; one-handed reach assumed. - **Screen 8 — review & export:** Web (Epic embed) for QA nurses, charge nurses, and Joint Commission auditors. Wider transcript view, keyboard navigation, multi-select for batch export. - **Telehealth overlay (future):** Web. Joins Zoom / Teams / Epic telehealth as a live-caption participant. ### Why not React Native / Flutter for v1 - Live audio + mic-state visuals need tight OS integration. Cross-platform abstractions add a frame of latency per audio chunk that hurts the perception of real-time. - HIPAA + clinical-safety review is easier when the audit boundary is one platform per submission. - The two launch platforms (iOS + Web) genuinely share little UI code anyway — the iPhone bedside layout and the Epic-embedded review console are different products in the same family. ### Design-system implications - **iOS:** Apple HIG, SF Symbols, SF Pro / SF Pro Rounded for patient-facing copy (rounder reads less clinical). Dynamic Type up to AX5. - **iPad bedside:** Large-touch HIG variant, WCAG 2.2 AA contrast minimum, two-finger drag to swap clinician/patient halves of the screen. - **Web:** Tailwind + Radix UI on Next.js. Design tokens (color, type scale, spacing) shared with iOS via a JSON token file generated from Figma variables. - **Right-to-left:** First-class from day one — Arabic is in the launch language set. All layouts use logical properties (start/end) not directional (left/right). - **Accessibility:** VoiceOver / TalkBack first-class. Dynamic Type / browser zoom honored. Tap targets ≥ 44pt clinician side, ≥ 56pt patient side (see [§5](#5-important-states--interactions--cross-cutting)). - **Localization:** All UI chrome is translatable; the 6 launch languages are also UI-chrome languages — a Spanish-fluent patient sees the consent screen *and* any chrome they touch in Spanish. ### What this rules out for v1 - **Glassmorphism / heavy translucency** — accessibility risk under fluorescent hospital lighting. - **Skeuomorphic shapes** — clinical interfaces value information density over delight. - **Dark mode only** — high-contrast light mode first (hospital lighting is bright); dark mode is a fast-follow, not a launch deliverable. - **Camera-based features** — no scanning prescriptions, no reading wristbands. Privacy footprint stays small. - **Cross-clinician-language combos** — English-only clinician side for v1 (a Spanish-native nurse talking to a Mandarin-native patient ships later). --- ## 3. User Journey Snapshot (Bedside) **Persona:** Maya, ED nurse, 6-year tenure. Mid-shift. Patient room 14 just admitted — Spanish-speaking, abdominal pain, no family present. The on-call interpreter line is showing a 12-minute queue. | # | Step | Time | Maya's POV | Patient's POV | |---|---|---|---|---| | 1 | Launch from chart | 0:00 | Tap "Translator" inside Epic on her phone | – | | 2 | Pre-flight | 0:03 | Sees patient name, "Spanish" pre-filled from chart, taps Start | – | | 3 | Consent | 0:05 | Hands phone to patient, points at screen | Hears + sees consent in Spanish, taps "Sí, entiendo" | | 4 | Live conversation | 0:10 | Speaks naturally, watches translation appear for patient | Hears Spanish, sees Spanish caption, replies — Maya sees English | | 5 | Clarity flag | 0:45 | "Discharge" flagged → taps "leave hospital" | Patient hears the unambiguous version | | 6 | Patient distress | 1:30 | Patient agitated, Maya taps "Get a human now" | Live interpreter joins in 11s | | 7 | End + save | 5:00 | Taps End → reviews 6-line transcript → Save to chart | Hands phone back | | 8 | (Later) Audit | – | Charge nurse reviews transcript next morning, exports PDF for QA | – | **Goal time-to-first-translated-word:** under 15 seconds from chart-tap. (Beats Jeenie's 13s human-connect — see [`RESEARCH.md` §1](RESEARCH.md).) --- ## 4. Key Screens Eight screens cover the hero flow + 2 supporting flows. Each one has a purpose, a wireframe sketch (ASCII), and the states it must handle. --- ### Screen 1 — Home / Today's encounters **Purpose:** Clinician landing page. Surfaces patients whose chart-listed language ≠ clinician's language. ``` ┌──────────────────────────────────────┐ │ 🔒 Encrypted · Maya R., RN · ED │ ├──────────────────────────────────────┤ │ Today [⚙] │ │ │ │ ⚠ NEEDS LANGUAGE SUPPORT (3) │ │ ┌──────────────────────────────────┐ │ │ │ Rm 14 · Ortiz, J. Spanish ▶ │ │ │ │ Rm 22 · Chen, L. Mandarin ▶ │ │ │ │ Rm 31 · Khalil, A. Arabic ▶ │ │ │ └──────────────────────────────────┘ │ │ │ │ Recent sessions (7) ▶ │ │ Phrasebook (offline) ▶ │ │ │ │ ┌──────────────────────────────────┐ │ │ │ + Start translator │ │ │ └──────────────────────────────────┘ │ └──────────────────────────────────────┘ ``` - **Encryption chrome** is persistent at the top — every screen, never just splash. - **"Needs language support" list** comes from EHR's language-preference field. Most-common failure in clinic: nurse forgets it's on file. Surfacing it removes that step. - **Recent sessions** is the entry point to Screen 8. - **Offline phrasebook** is a fallback when network drops (Care to Translate pattern — see [`RESEARCH.md` §2](RESEARCH.md)). **States:** loaded, loading, empty ("No language-flagged patients today"), offline (yellow banner). --- ### Screen 2 — Pre-flight / Encounter setup **Purpose:** 3-second confirmation before the conversation starts. ``` ┌──────────────────────────────────────┐ │ ← Setup │ ├──────────────────────────────────────┤ │ │ │ PATIENT │ │ Jorge Ortiz · MRN 4827193 │ │ Room 14 · ED · Encounter #3 │ │ │ │ LANGUAGE PAIR │ │ ┌────────────────────────────────┐ │ │ │ English ⇄ Español [change] │ │ │ └────────────────────────────────┘ │ │ ⓘ Auto-detected from chart │ │ │ │ ENCOUNTER TYPE │ │ [Triage] [Bedside ✓] [Discharge] │ │ ⓘ Loads relevant verified phrases │ │ │ │ ☑ Save transcript to chart on end │ │ ⓘ Otherwise deleted at session end │ │ │ │ ┌────────────────────────────────┐ │ │ │ → Hand phone to patient │ │ │ └────────────────────────────────┘ │ └──────────────────────────────────────┘ ``` - **Auto-population from EHR** (Voyce/LanguageLine pattern — [`RESEARCH.md` §2](RESEARCH.md)) means most encounters need zero edits here. - **Encounter type** is the lever for Clinical Clarity — different verified-phrase libraries load per type. - **Retention checkbox** makes the "not stored unless saved" promise tangible. Default *on* for clinical; *off* for casual. **States:** auto-filled, manual-override, no-EHR-data warning, offline (verified phrases only). --- ### Screen 3 — Patient consent (patient-facing) **Purpose:** Earn trust in 5 seconds, in the patient's language, with audio for low literacy. ``` ┌──────────────────────────────────────┐ │ 🔒 │ │ │ │ Hola Jorge. │ │ │ │ Vamos a usar un traductor de IA │ │ para hablar con usted. │ │ │ │ ▶ Escuchar · Leer en voz alta │ │ │ │ • La traducción es en tiempo real │ │ • Sus palabras NO se guardan │ │ a menos que el enfermero lo pida │ │ • Puede pedir un intérprete humano │ │ en cualquier momento │ │ │ │ ¿Está de acuerdo? │ │ │ │ ┌────────────────┐ ┌──────────────┐ │ │ │ Sí, entiendo │ │ Necesito un │ │ │ │ │ │ humano │ │ │ └────────────────┘ └──────────────┘ │ └──────────────────────────────────────┘ ``` - **Patient's name first** — pulled from chart, never "Patient". - **Plain language, big type, calm.** No legalese. Audio playback because literacy ≠ language fluency. - **Three promises** map 1:1 to the three biggest fears: surveillance, AI inaccuracy, no-human-available. - **"I need a human" is a peer button**, not buried. Patient can opt out at any moment. This is the trust unlock. **States:** waiting for tap, audio playing (waveform animation), declined (routes to human queue), language not supported (falls back to top-6 picker with country flags — Martti's "point to flag" pattern). --- ### Screen 4 — Live conversation (face-to-face split) — **the hero screen** **Purpose:** Phone lies flat between Maya and Jorge. Each half oriented toward one person; clinician's captions inverted on top. ``` ┌──────────────────────────────────────┐ │ CLINICIAN VIEW (top) │ │ │ │ "Where does it hurt the most?" │ │ ──────────────────────────── │ │ tap to repeat · ⚠ flag │ │ │ │ Patient said: │ │ "Aquí, abajo del lado derecho." │ │ → "Here, on the lower right side." │ │ │ │ Quick replies: │ │ [On a scale 1–10?] [How long?] │ │ [Sharp or dull?] [Type custom] │ │ │ ├══════════════════════════════════════┤ │ 🎙 Listening · 🔒 Encrypted · ES⇄EN │ │ [⏸] [👤 Human] [⏹] │ ├══════════════════════════════════════┤ │ PATIENT VIEW (bottom) │ │ │ │ "¿Dónde le duele más?" │ │ │ │ ─── Su turno para hablar ─── │ │ │ │ ▼ Toque para hablar │ │ │ └──────────────────────────────────────┘ ``` - **Inverted top half** = clinician's view when phone is flat. Patient sees their language right-side-up across the table. - **Translation dominates; original is secondary caption.** Clinician can audit; patient can read distraction-free. - **Mic state** (`🎙 Listening`) lives in the center stripe, glanceable from a meter away. - **Quick replies** on clinician side only — patient side stays simple. (Pattern: Google Interpreter Mode, [`RESEARCH.md` §2](RESEARCH.md).) - **Three center controls:** pause, escalate, end. Thumb-reachable for both speakers. - **No EHR clutter.** Chart is one tap away (`⏸` then back) but invisible mid-conversation. **States:** - `Listening` (mic open, animated) - `Translating` (brief shimmer, ~300ms) - `Speaking` (audio playing — the receiving half pulses) - `Paused` (both halves dim, big "Resume" button) - `Low confidence` (yellow border + Clarity flag → Screen 5) - `Network lost` (red banner: "Offline — verified phrases only") - `Battery <10%` (warning chip — shared shift devices die mid-conversation otherwise) --- ### Screen 5 — Clarity Filter disambiguation **Purpose:** When AI confidence on a medical term drops below threshold, **stop, force a choice, never silently guess.** Defensible differentiator vs. DeepL / Google. Appears as sheet over clinician half only. Patient half shows calm "one moment" message in their language. ``` ┌──────────────────────────────────────┐ │ ⚠ Ambiguous term │ ├──────────────────────────────────────┤ │ │ │ You said: │ │ "We'll discharge you tomorrow." │ │ │ │ 'discharge' has 3 meanings in │ │ Spanish. Which one? │ │ │ │ ┌──────────────────────────────────┐ │ │ │ ● Leave the hospital │ │ │ │ (dar de alta) │ │ │ ├──────────────────────────────────┤ │ │ │ ○ Bodily fluid │ │ │ │ (secreción) │ │ │ ├──────────────────────────────────┤ │ │ │ ○ Electrical │ │ │ │ (descarga) │ │ │ └──────────────────────────────────┘ │ │ │ │ ☐ Always use this for me │ │ │ │ [Cancel translation] [Confirm →] │ └──────────────────────────────────────┘ ``` - **Original phrase quoted back** — clinician can correct themselves rather than just pick. - **Plain-English gloss + verified target phrase** under each option. The verified phrase signals this isn't a guess. - **"Always use this for me"** is personalization — Maya picks "leave the hospital" 99% of the time; the flag fades for her. - **"Cancel translation"** is the safety valve — if none fit, she rephrases instead of forcing a wrong one through. **States:** triggered, expanded definition (tap option), patient-side "one moment", abandoned ("Let me rephrase…" to patient). --- ### Screen 6 — Escalate to human **Purpose:** One-tap fallback when AI isn't trustable enough — patient distress, dialect mismatch, end-of-life, complex consent. ``` ┌──────────────────────────────────────┐ │ ← Connecting a human interpreter │ ├──────────────────────────────────────┤ │ │ │ 🎧 Spanish · Medical certified │ │ │ │ ◐ Searching… │ │ Average wait: 11 seconds │ │ │ │ Position in queue: 1 │ │ │ │ Meanwhile, AI translation stays │ │ on. You can keep talking. │ │ │ │ ───────────────────────────── │ │ │ │ While you wait: │ │ ┌────────────────────────────────┐ │ │ │ [En español] Espere un momento │ │ │ │ por favor — un intérprete │ │ │ │ humano está conectando. │ │ │ │ ▶ Reproducir para el paciente │ │ │ └────────────────────────────────┘ │ │ │ │ [Cancel — keep using AI] │ └──────────────────────────────────────┘ ``` - **AI doesn't stop while waiting** — conversation continues; human joins on top. Removes the "everything freezes" anxiety. - **Pre-recorded reassurance phrase** for the patient. One tap plays it. - **Honest wait time** — pulled from real queue depth, not marketing. - **Cancel always available** — opting in to a human is not a one-way door. **States:** searching, connected (Screen 4 transitions to 3-way with interpreter tile), no-one-available ("All Spanish interpreters busy — try phone? [Y/N]"), cancelled. --- ### Screen 7 — Session end / save to chart **Purpose:** 10-second decision: save, edit, or discard. Defaults make the safe choice automatic. ``` ┌──────────────────────────────────────┐ │ Session ended · 4 min 52 sec │ ├──────────────────────────────────────┤ │ │ │ Jorge Ortiz · Spanish · Bedside │ │ 12 exchanges · 2 clarity flags │ │ 0 human-interpreter minutes │ │ │ │ ───────────────────────────── │ │ │ │ TRANSCRIPT (12) │ │ ┌──────────────────────────────────┐ │ │ │ Maya: Where does it hurt most? │ │ │ │ Jorge: Here, lower right side. │ │ │ │ ⚠ "discharge" → "leave hospital"│ │ │ │ Maya: We'll discharge you tom… │ │ │ │ ... [tap to expand] │ │ │ └──────────────────────────────────┘ │ │ │ │ ┌──────────────────────────────────┐ │ │ │ Save to chart │ │ │ └──────────────────────────────────┘ │ │ Discard Edit transcript │ └──────────────────────────────────────┘ ``` - **Numbers tell a story:** 12 exchanges, 2 clarity flags, 0 human minutes — a charge nurse can audit in 3 seconds. - **Clarity flags inline in transcript** — audit trail shows *which* terms were disambiguated. Defensible in malpractice review. - **Save is primary.** Discard secondary — defaulting to save reduces "did I lose it?" anxiety. Patient already consented on Screen 2. - **Edit transcript** = per-line view, edits tracked, never silent overwrites. **States:** default, saving (spinner → "Saved to Epic ✓"), error (retry / save locally), discarded ("Transcript deleted ✓ · Audit log retained"). --- ### Screen 8 — Session history & review/export (supporting flow) **Purpose:** Compliance lives here. QA nurse, malpractice reviewer, Joint Commission auditor — all walk in cold and find what they need. ``` ┌──────────────────────────────────────┐ │ Sessions · Today ⌕ │ ├──────────────────────────────────────┤ │ ┌──────────────────────────────────┐ │ │ │ 14:32 Ortiz, J. Spanish │ │ │ │ Bedside · 4:52 · ⚠ 2 · Maya R. │ │ │ │ [Saved to chart ✓] │ │ │ └──────────────────────────────────┘ │ │ ┌──────────────────────────────────┐ │ │ │ 11:08 Chen, L. Mandarin │ │ │ │ Triage · 2:14 · ⚠ 0 · Maya R. │ │ │ │ [Discarded — audit log only] │ │ │ └──────────────────────────────────┘ │ │ │ │ Filters: [Date] [Lang] [Clinician] │ │ [⚠ Clarity flags] [Human] │ │ │ │ Export selected (1) → [PDF] [CSV] │ └──────────────────────────────────────┘ ``` Tap into a session → ``` ┌──────────────────────────────────────┐ │ ← Ortiz, J. · 14:32 · Spanish │ ├──────────────────────────────────────┤ │ AUDIT TRAIL │ │ ✓ Encrypted (AES-256) │ │ ✓ Consent captured 14:32:05 │ │ ⚠ 2 Clarity flags resolved │ │ ✓ Saved to Epic by Maya R. │ │ ✓ Hash: 4f3a…b921 │ │ │ │ TRANSCRIPT [Expand all] │ │ 14:32:11 Maya: Where does it… │ │ 14:32:14 Jorge: Aquí, abajo… │ │ 14:32:18 ⚠ "discharge" → "leave │ │ hospital" — Maya │ │ ... │ │ │ │ [Flag for clinical review] │ │ [Export PDF] [Send to Epic] │ └──────────────────────────────────────┘ ``` - **Hash on audit trail** = tamper-evidence. If the PDF in court doesn't match the hash, someone edited it. - **Filter by Clarity flags / Human escalation** lets QA find high-risk sessions first. - **"Flag for clinical review"** = one-tap route to the clinical safety team. Closes the loop on edge cases. **States:** list, filtered, empty, individual session, export-in-progress, exported (PDF preview). --- ### Bonus — Intake kiosk variant (supporting flow) Same engine, different framing. Single device on a stand at registration; patient self-serves before being seen. ``` ┌──────────────────────────────────────┐ │ Welcome · 你好 · Bienvenido · أهلاً │ │ │ │ Tap your language / flag │ │ │ │ 🇺🇸 🇲🇽 🇨🇳 🇻🇳 🇸🇦 🇭🇹 │ │ EN ES ZH VI AR HT │ │ │ │ ──── or ──── │ │ 🎙 Just start talking │ │ We'll detect your language │ └──────────────────────────────────────┘ ``` After language pick → structured Q&A: ``` ┌──────────────────────────────────────┐ │ 1 of 6 · Español │ ├──────────────────────────────────────┤ │ │ │ ¿Qué le trae hoy? │ │ │ │ ▼ Toque para responder │ │ ──────────────────────── │ │ │ │ O elija: │ │ [Dolor] [Fiebre] [Lesión] │ │ [Cita] [Receta] [Otro] │ │ │ │ [← Atrás] [Saltar pregunta →] │ └──────────────────────────────────────┘ ``` - **Flag picker + voice fallback** = Martti's "point to country" + Google's auto-detect. Belt-and-suspenders for low-literacy patients. - **Chip answers + freeform voice** — patients who freeze at open-ended questions can pick a chip. - **Skip allowed.** Never force an answer; flags blank fields for nurse follow-up. - **Output:** structured intake note auto-saved to chart in English, transcript attached. Nurse opens chart and sees the patient's answers already typed up. --- ## 5. Important States & Interactions — Cross-Cutting Persistent across every screen: | State | Where it shows | Why | |---|---|---| | Encryption / "not stored" badge | Top status bar, every screen | Trust signal that doesn't have to be re-earned per screen ([`RESEARCH.md` §4](RESEARCH.md)) | | Language pair pill | Center stripe on Screen 4, top on others | Always 1-tap-swappable; never have to back out | | Network state | Top banner when degraded | Drops to verified phrasebook gracefully; never silent-fail | | Mic state | Animated icon, center stripe | Glanceable across the room | | Battery low | Small chip when <10% | Shared shift devices die mid-conversation otherwise | | Audit-log tick | Bottom-right, faint | Every interaction logged; reassures without dominating | **Interaction rules:** - **No silent failures.** Network drop → user-visible banner. Low confidence → Clarity flag. Mic permission denied → explicit fix-it dialog. - **Patient half is always calmer.** Bigger type, fewer controls, more whitespace, slower animations. Patient is anxious; clinician is competent. - **One-tap escalation everywhere.** "Get a human" reachable from Screens 2, 4, 5 — never more than one tap away mid-conversation. - **Defaults are safe.** Save to chart on, retention on, encryption on. Clinician opts out, not in. - **Tap targets ≥ 44pt** even on patient side; older patients with shaky hands shouldn't have to aim. --- ## 6. Assumptions Made Restated from [`RESEARCH.md` §6](RESEARCH.md) so reviewers can challenge each: 1. **Single device, not paired** — one phone or tablet between clinician and patient (Martti/DeepL precedent; cheaper to deploy). 2. **Clinician is the primary authenticated user.** Patient interacts through a "patient mode" the clinician hands them — chart already identifies them. 3. **AI-first with human fallback**, not human-first with AI as backup. AI confidence is good enough to default; human is the escalation. Clinical Clarity Filter makes this defensible. 4. **EHR integration is post-MVP** for this submission, but every screen anticipates it — Screen 1 reads language from EHR, Screen 7 writes transcript back, Screen 8 surfaces Epic save status. 5. **"Clinical Clarity Filter"** = real-time confidence scoring with mandatory disambiguation below threshold. Threshold tunable per encounter type (stricter for Discharge / Med rec). 6. **Six launch languages:** Spanish, Mandarin, Vietnamese, Arabic, Haitian Creole, Russian — six largest US LEP populations. Auto-detect across all. 7. **Two clinician personas at launch:** ED nurse (Maya, hero) + Registration clerk (kiosk). Doctors and discharge planners are out of scope for this submission. 8. **HIPAA-ready posture, not certified.** BAA-able; SOC 2 Type II target inside 12 months; on-device processing where feasible. 9. **No video on AI side.** AI is voice + text only. Video joins only when a human interpreter escalates. Deliberate scope cut to keep the bedside footprint tiny. 10. **English is the clinician language for v1.** Other clinician-language combos are post-MVP — combinatorial complexity not justified yet. --- ## 7. What's Deliberately Not Designed Yet - **Admin console** for hospital IT — user provisioning, retention policy, language-pack management. - **Discharge / med-rec flow** — high stakes; deserves its own pass with pharmacist input. - **Telehealth 3-way overlay** — covered in [`RESEARCH.md` §5C](RESEARCH.md); DeepL/Voyce already in this space, less novel. - **Onboarding for first-time clinicians** — assumed they've done a 5-minute training video. - **Multi-tenant white-labeling** — hospitals will want this; system anticipates but not specced. --- ## 8. Open Questions for the Stakeholder 1. **Patient consent: tap or signature?** Tap is faster; signature more legally defensible. Likely state-by-state. 2. **Clarity threshold: clinician setting or admin setting?** Lean admin (consistency); clinicians will want override. 3. **Audit log retention?** Default 6 years (HIPAA minimum), configurable per facility. 4. **Human escalation cost visible to clinician?** Hidden cost may create over-use; visible cost creates reluctance. Recommend silent. 5. **Offline behavior on IV-cart iPad?** Verified phrases only, or verified + on-device small model? Speed vs accuracy tradeoff. --- ## 9. Next Iterations | Pass | Goal | |---|---| | **Mobbin reference** | Pull 10–20 real bedside / kiosk / consent screens once `mcp__mobbin__*` is in-session (after restart). Validate this flow against known anti-patterns. | | **Clinician validation** | Walk the flow with one ED nurse + one registration clerk. Test: pre-flight time, consent comprehension, hand-off moment. | | **Hi-fi mockups** | Translate ASCII wireframes to Figma. Apply trust-signal chrome from [`RESEARCH.md` §4](RESEARCH.md) consistently. | | **Prototype** | Clickable prototype of Screen 4 with simulated mic states and a forced Clarity flag — the moment that sells the differentiator. | | **Pilot scope** | One ED, Spanish, 30 days. Measure: time-to-first-word, # Clarity flags resolved, # human escalations, transcript-save rate. |