dyolink/docs/specs/voice-treatment-entry/spec.md

Voice treatment entry

Status: Implemented on feat/voice-treatment-entry. First live test on 2026-08-21 sent the tooth path back for revision — a spoken number is now read as its FDI code (§6). Still blocked on the ASR spike (§11 item 1) before it is trustworthy in front of patients Area: Treatment workspace (CLINIC orgs) Created: 2026-08-20

Fill a TreatmentDetail — including its lab dispatch — by speaking, instead of by tapping through the type dropdown, the FDI chart, the prosthesis wizard and the lab picker.

---

1. Goal

A clinician on the Treatment tab taps a microphone, describes the treatment for the already-selected patient in one utterance, and is shown a review sheet of what was understood. Fields they tick are applied to the open detail chip. Nothing is written to the form without confirmation.

In scope

One recording produces exactly one TreatmentDetail, and may fill every field of it:

Field	Source
treatmentType	catalog code, matched against locale labels
teeth	FDI codes, via tooth-intent resolver
toothSelectionGroups	connected (bridge) / single spans
comment	cleaned dictated notes
lab: prosthesisTypeCode per tooth	default type + per-tooth overrides
lab: destinationOrganizationId	matched against the clinic's linked labs
lab: dueDate	via due-date intent resolver

Out of scope (v1)

• Multiple detail chips from one recording.
• attachmentIds — files cannot be dictated.
• Editing an existing detail by voice ("no, make that 15"). Confirming a recording always creates a new detail (see §2).
• Creating the treatment or selecting the patient by voice. A patient is already selected; voice only fills the form.
• Lab-side (LAB org) usage. Clinic only.

---

2. User flow and UI integration

The control: Add detail, split

The Add detail button gains a second segment holding the mic. The halves read as siblings — both end in a new detail — but they are independent actions:

• Add half — unchanged. Same onAddDetail, same seeding, same setEntryStep. It gains a neighbour and nothing else. Its logic is not modified, wrapped or made conditional.
• Mic half — starts a recording. Nothing is created until confirm (below).

The Add detail <Button> in TreatmentDetailsEditor becomes a segmented control built exactly like the detail chip's trash affordance in the same file (TreatmentDetailsEditor.tsx, the chip <div> + label <button> + remove <button>):

• an inline-flex items-stretch overflow-hidden rounded-[var(--radius-md)] wrapper;
• two raw <button> children divided by border-s, each with its own disabled, title, aria-label and focus-visible:ring-inset;
• styled primary (bg-primary text-white) to preserve the button's current look;
• w-full sm:w-auto on the wrapper with the Add half flex-1, reproducing today's fullWidth + sm:w-auto shrink-0 behaviour.

The shared Button component is not reusable for the halves: it hardcodes rounded-[var(--radius-md)] on each instance and owns auto-pending state, both of which fight a segmented control. This is precisely why the chip pattern uses raw <button>s, and this control follows it.

Side: the mic is the second flex child with border-s — the logical end, exactly like the trash. Visually right in en/nl, visually left in fa. No physical left/right anywhere, per the repo's RTL rule.

en / nl (LTR)          fa (RTL)
[  Add detail │ 🎤 ]   [ 🎤 │  افزودن  ]
[ ترمیم 14,15 │ 🗑 ]   [ 🗑 │ ترمیم 14,15 ]
        ↑ same side as the chip's trash, in both directions

Flow

[editable day]
        │
        ▼  tap 🎤        ← mic swaps to ■ (red); Add half disabled
   recording ─────────── inline bar below the header row:
        │                ● 0:12 / 2:00  ▁▃▇▅▂▆█▄▁   [لغو]
        │                auto-stops at the 2:00 cap
        ▼  tap ■
   processing ────────── ● transcribing…   ○ extracting     [cancel]
        │
        ▼
   review sheet
     ☑ Type        ترمیم
     ☑ Teeth       [mini FDI chart]  14 15
     ☑ Notes       حساسیت به سرما
     ☐ Lab         لابراتوار سینا      ⚠ similar name
        │
        ▼  [Apply n fields]
   detail created or filled → normal save flow

No layout shift. The segmented control never changes size; only the mic's icon and colour change. The timer and level meter live in a full-width bar inserted between the header row and the chip strip — the header is sm:justify-between, so growing the button mid-recording would shove the row on every start and every stop.

Duration is capped at 2 minutes (maxMs, configurable, v1 default 120_000). The timer shows elapsed / 2:00 and the recorder auto-stops at the cap. maxMs: null means uncapped and remains supported, but is not the v1 default.

Two minutes is generous against the longest realistic utterance — a full prosthesis dictation with type, several teeth, a bridge, prosthesis type, lab and due date — while bounding worst-case spend. It also sits comfortably inside the vendor limits (§4): 2 minutes of webm/opus is well under 1 MB against a 25 MB ceiling, and transcribes in a few seconds against a 60s processing timeout.

Cost per recording is therefore bounded at $0.012 of ASR (2 min × $0.006) plus ~$0.0008 of extraction — about 1.3¢ worst case, against ~0.3¢ for a typical 20s utterance.

What confirm does

The detail is created on confirm, never on tap. Tapping the mic starts a recording and nothing else, so cancelling, a vendor failure, a rate-limit, or navigating away leaves the chip strip untouched — there is no orphan state to clean up.

Confirm always appends a new detail. One unconditional rule, no dependence on invisible state: append newDetail() seeded with defaultTreatmentTypeForAppointment, make it active, setEntryStep('treatment'), then apply the ticked rows on top — so an extracted type overrides the seed, and unticking the type row leaves the seeded default in place.

This reuses newDetail() and the same state transitions, but it is a separate code path. onAddDetail is not called and not changed.

setEntryStep('treatment') matters: showChrome is always on, so the control is visible during the Lab wizard step too. Confirming there returns to the treatment step.

Accepted consequences:

• Tapping Add and then 🎤 leaves behind the blank chip that Add created. It carries the usual trash affordance.
• Dictating into an existing detail is not supported in v1 — voice always makes a new one.

Render policy

Three different reasons for "no", rendered differently:

Reason	Condition	Render
Technical	no voice profile for the locale; no MediaRecorder	segment absent — the control is byte-for-byte today's plain Add button
Commercial	Plan.features.voiceTreatmentEntry false	segment absent — ⚠ not enforced in v1, see §8
Contextual	!canEdit || disabled — past day, read-only load, no TAB_TREATMENT_EDIT	both segments render and disable together, like the chip's trash (disabled:opacity-40 disabled:cursor-not-allowed)

"This feature isn't yours" and "not right now" are different statements. Absence avoids a permanently dead control; disabling avoids the button resizing as the day strip moves.

TAB_TREATMENT_EDIT is resolved via common/membership-permissions.ts, never by reading membership.permissions directly.

Component API

TreatmentDetailsEditor gains exactly one optional prop. All behaviour — MediaRecorder, the API call, error state — lives in lib/voice/useVoiceCapture.ts (hooks belong in lib/ per AGENTS.md) and is owned by TreatmentWorkspace. The editor stays presentational and renders both the segment and the recording bar with its own classes, keeping the segmented styling beside the chip pattern it mirrors.

/** Omit when voice is unavailable — the Add button then renders unsplit. */
voice?: {
  phase: 'idle' | 'recording' | 'processing';
  elapsedMs: number;
  level: number;   // 0..1, for the meter
  maxMs: number | null;   // v1: 120_000 (2 min). null = uncapped, supported but not default
  onStart: () => void;
  onStop: () => void;
  onCancel: () => void;
};

voice === undefined is the absent state above — availability is expressed by presence rather than a separate flag, so the two cannot disagree.

Accessibility

• Mic segment carries both title and aria-label, like the chip's trash, and its label changes with phase.
• Announce phase transitions via a role="status" aria-live="polite" region — the same idiom as the existing autosave status line. Do not put aria-live on the ticking timer.
• Focus ring is focus-visible:ring-inset tinted primary when idle, red while recording.

---

3. Architecture

Two sequential stages, both server-side. The vendor API key never reaches the browser.

browser ──audio(base64)──► POST /treatments/voice-extract
                                    │
                          ┌─────────┴─────────┐
                          │ 1. AsrProvider    │  audio + locale hint → transcript
                          └─────────┬─────────┘
                          ┌─────────┴─────────┐
                          │ 2. ExtractionProv │  transcript + catalog + ctx → VoiceIntent
                          └─────────┬─────────┘
                          ┌─────────┴─────────┐
                          │ 3. resolvers      │  intents → FDI codes, ISO date
                          │    (pure, tested) │
                          └─────────┬─────────┘
                                    ▼
                          VoiceExtractionResult  (resolved values + the intents
                                                  that produced them, for display)

Why intents and not final values: the model never emits an FDI code and never does calendar arithmetic. It emits what it heard; deterministic, unit-tested code decides what that means. This is what makes the two highest-consequence mappings — quadrant mirroring and Jalali conversion — testable instead of hopeful.

Endpoint

POST /treatments/voice-extract

• Guards: JwtAuthGuard + ClinicOrgGuard.
• Service-level check of TAB_TREATMENT_EDIT. The plan flag is not checked in v1 (§8).
• Per-user throttle via @nestjs/throttler (present in package.json, currently wired nowhere in src/ — this is its first use, so the module must be registered in app.module.ts). Configurable; v1 default 6 requests / 60s per user (VOICE_THROTTLE_LIMIT, VOICE_THROTTLE_TTL). A human cannot approach that — a recording plus processing takes ten seconds at minimum — so it is purely an abuse and runaway-loop guard, which is doing more work than usual given v1 is ungated and uncapped (open item 14).
• Body: base64 audio + declared format + clientTodayIso + IANA timeZone + treatmentDetailId/clientId for context.

clientTodayIso and timeZone come from the client per the existing house rule (AGENTS.md): never derive the clinic's civil day from Date#getDay() on the UTC server. Reuse common/zoned-civil-time.ts.

---

4. Provider registry

ASR and extraction are separate swappable roles. They will not come from the same vendor for every locale.

interface AsrProvider {
  transcribe(audio: AudioInput, localeHint: string): Promise<{ text: string; usage: AsrUsage }>;
}

interface ExtractionProvider {
  extract(transcript: string, catalog: CatalogPrompt, ctx: ExtractionContext): Promise<VoiceIntent>;
}

Resolved through a registry keyed by locale.

At launch all three locales use the same profile — OpenRouter with openai/whisper-1 for ASR. The per-locale indirection is kept anyway, because the locale is the axis along which this is most likely to diverge: Persian ASR is the weakest link (§11), and swapping only fa to a Persian-specialist vendor must not be a code change.

VOICE_PROFILE_FA = openrouter:openai/whisper-1 | openrouter:<llm-model>
VOICE_PROFILE_EN = openrouter:openai/whisper-1 | openrouter:<llm-model>
VOICE_PROFILE_NL = openrouter:openai/whisper-1 | openrouter:<llm-model>

The language hint is the profile's locale as ISO-639-1 — fa, en, nl — not a constant.

Reachability is a property of where you deploy, not of the code. The same image serves an Iran-hosted instance with a domestic fa profile and a Europe-hosted instance with an OpenRouter profile; only config differs. A locale with no configured profile has no microphone button at all (§2).

The frontend must learn which locales are enabled from the API, not from a NEXT_PUBLIC_* var — those are baked in at build time, so an env-var approach would make enabling a locale require rebuilding and repushing the frontend image.

v1 provider: OpenRouter

Both roles, one API key.

ASR — POST https://openrouter.ai/api/v1/audio/transcriptions

{
  "model": "openai/whisper-1",
  "input_audio": { "data": "<base64>", "format": "webm" },
  "language": "fa"
}

• Accepted formats: WAV, MP3, FLAC, M4A, OGG, WebM, AAC. Chrome/Android webm/opus and Safari/iPad mp4/aac both go through unmodified — no transcode dependency is required.
• Limits: 25 MB; 60s upstream processing timeout. The 2-minute recording cap (§2) sits comfortably inside both.
• Request shape verified against OpenRouter's STT docs: base64 JSON input_audio is the documented primary path (multipart file is the OpenAI-compatible alternative), and both default model slugs exist on the live models API.
• The cap is enforced twice: the client's reported durationMs, and again against the vendor's own usage.seconds — the client's figure is a claim, not enforcement. The server allows a 2s tolerance, because the client measures length after the recorder stops and a recording that runs to the cap always reports slightly over it.
• Response: { text, usage: { seconds, total_tokens, cost } }.
• Price: openai/whisper-1 is $0.006/minute, billed to the nearest second → $0.002 for a typical 20s utterance, $0.012 at the 2-minute cap. OpenRouter forwards this model directly to OpenAI.

Extraction — OpenRouter chat completions with a JSON schema constraining the output to VoiceIntent. Model id is config (VOICE_LLM_MODEL).

Chosen: google/gemini-3.7-flash.

Cost is deliberately not the selection axis. The call is ~1,250 input tokens (system prompt + catalog labels + JSON schema + linked labs + transcript) and ~200 output, so the whole candidate field spans about one cent per recording:

Model	in / out per M tokens	≈ per call
google/gemini-3.7-flash	$0.375 / $1.875	$0.0008
qwen/qwen3.8-27b	$0.45 / $3.20	$0.0012
qwen/qwen3.8-max	$2 / $6	$0.0037
openai/gpt-5.6-terra	$2 / $12	$0.0049
anthropic/claude-opus-5	$5 / $25	$0.0113

Select for Persian comprehension and reliable constrained JSON instead. Gemini Flash wins on the reasoning that this task is not reasoning-heavy — read a short sentence, pick codes from a supplied closed list, emit small JSON — and it is simultaneously the cheapest and lowest-latency candidate, which matters on a pipeline already at 3–10s.

Escalation path if Persian comprehension proves weak on real transcripts (item 1): qwen/qwen3.8-max (leads current multilingual rankings, ~4.6× the cost) and then anthropic/claude-opus-5 as the accuracy ceiling. Both are a config change — VOICE_LLM_MODEL — not a code change. Worth knowing when judging that ranking evidence: it is marked provisional and was measured on the previous Qwen generation.

---

5. Extraction contract

The model returns intents only. Illustrative shape:

type VoiceIntent = {
  treatmentType: string | null;          // catalog code, from the supplied closed list
  teeth: ToothIntent[];
  connectedSpans: { from: ToothIntent; to: ToothIntent }[];
  comment: string | null;
  prosthesis: {
    defaultType: string | null;          // catalog code
    overrides: { tooth: ToothIntent; type: string }[];
  } | null;
  labId: string | null;                  // must be one of the supplied linked-lab ids
  labMatchExact: boolean;
  due: DueIntent | null;
};

type ToothIntent =
  | { kind: 'explicit'; fdi: string; spoken: string }
  | { kind: 'positional'; arch: 'upper' | 'lower';
      side: 'patient_right' | 'patient_left'; position: number; spoken: string };

type DueIntent =
  | { kind: 'weekday'; weekday: string; which: 'this' | 'next' }
  | { kind: 'offset'; unit: 'day' | 'week' | 'month'; amount: number }
  | { kind: 'jalali'; jy: number; jm: number; jd: number }
  | { kind: 'gregorian'; y: number; m: number; d: number };

Every code-valued field is constrained to a closed list supplied in the prompt:

• Treatment types and prosthesis types come from CatalogLabelService in the actor's locale, so the model sees "پروتز" and "زیرکونیا مونولیتیک" as the spoken forms of prosthesis and monolithic_zirconia rather than being asked to translate. Catalog entities store a stable code and no label — never hardcode a label.
• Lab candidates are the clinic's linked labs only (OrganizationLink), passed as { id, name }. The model may return one of those ids or null, nothing else.

Every unresolved or rejected item is reported, never silently dropped.

---

6. Resolvers

Both live in backend/src/, pure and Jest-covered. The frontend has no test runner (no jest/vitest, zero spec files) — putting them there would forfeit the testability that justified this whole design.

resolveToothIntent()

• Owns the patient-right convention in exactly one place: upper + patient_right → quadrant 1, upper + patient_left → 2, lower + patient_left → 3, lower + patient_right → 4. This is the mirroring bug, and it becomes a unit test.
• Rejects out-of-range positions rather than clamping. Position 9 is unresolved, never 8.
• Permanent dentition only — FDI 11–18/21–28/31–38/41–48, matching FDI_UPPER_LEFT_TO_RIGHT / FDI_LOWER_LEFT_TO_RIGHT. Deciduous references ("دندان شیری") must resolve to unresolved, never snap to a permanent tooth.
• Locale-neutral by construction. ToothIntent carries arch/side/position, not words, so the resolver needs no per-locale branches. The locale-specific part is the prompt: each enabled locale needs its own spoken tooth vocabulary (شش بالا راست, upper right six, rechtsboven zes).
• A spoken tooth number is an FDI code, in every locale. This is how clinicians actually dictate — "بیست و شش" is tooth 26 — so the prompt teaches the notation (first digit = quadrant from the patient's own point of view, second = position from the midline) rather than refusing it. arch/side/position is the reading of a tooth that was described instead of numbered, where a single digit is a position and the quadrant comes from words. Revised after the first live test; the original design had this backwards and made the descriptive form the only supported path.
• A single digit alone is never resolved. "دندون دو" names four teeth. It is reported as tooth_missing_quadrant with the candidate codes attached — narrowed by whatever was said, so "دو بالا" offers two — and the review sheet turns them into chips. The clinician chooses; the resolver still never guesses.
• Digits arrive in three scripts. normalizeFdiCode (common/fdi.ts) folds Persian and Arabic-Indic digits to ASCII and strips the spaces of a digit-by-digit dictation before anything is matched, at both the wire branch choice and the final validation.

resolveDueDate()

• Takes clientTodayIso + IANA timeZone; reuses common/zoned-civil-time.ts.
• Week start is per locale, because "next Thursday" is week-relative: fa starts Saturday, en and nl start Monday. Hardcoding Saturday put an en/nl clinician's deadline a week out. One place (weekStartForLocale), tested in both directions.
• 'this' is occurrence-anchored (soonest strictly-future, never resolves into the past); 'next' is week-anchored. A missing qualifier is read as 'this' — a bare weekday carries none, and failing would discard a real spoken deadline.
• Jalali conversion is arithmetic, not inference: port jalaliToGregorian and toLatinDigits from frontend/src/lib/i18n/persianCalendar.ts into backend/src/common/jalali.ts with a spec. It is dependency-free integer math (~60 relevant lines) and the calendar does not change, so the duplication is stable.

Group / prosthesis rules

• Connected spans validate through the shipped helpers — areArchNeighbors, sameArch, teethBetweenInclusive. Never a 1-tooth connected group. Anything invalid degrades to singles and is flagged on the review sheet.
• Prosthesis: expand defaultType across all teeth, then apply per-tooth overrides.
• All-or-nothing. assertCompleteToothProsthesisMap requires every tooth on a prosthesis detail to carry a prosthesisTypeCode or the send throws TREATMENT_TOOTH_PROSTHESIS_INCOMPLETE. So if even one tooth ends untyped, the prosthesis row is marked incomplete and stays unticked — the unshippable state surfaces at review, where it is cheap, not at dispatch minutes later on another screen.
• Apply order is teeth → groups → prosthesis, so pruneToothProsthesisForGroups behaves.

---

7. Review sheet

Presentation: a modal on desktop; a full-screen overlay on mobile.

⚠ On mobile this must be an overlay rendered in place — not a Next.js route. A real navigation unmounts TreatmentWorkspace and destroys the in-progress draft. This is the same constraint the realtime soft-refresh already lives under: never remount the treatment form, never clear a draft.

• Renders the transcript, then one row per extracted field in the app's own vocabulary: translated catalog labels, and a mini FDI chart for the teeth rather than a list of numbers.
• Each row has a checkbox. Ticked rows apply; nothing else is touched. Confirm is also what creates the new detail — see §2.
• Rows default to ticked except:
  • the lab row when labMatchExact is false — shipping to a lab always requires a deliberate tick;
  • any row carrying an unresolved item or an incomplete prosthesis map.
• Unresolved items are shown with what was heard ("دندان شیری — بازشناسی نشد"), so the clinician can see what the system did not understand.
• An item that carries candidates renders them as tappable chips — the one place the sheet is interactive. Picking one folds the tooth into the result (withChosenTeeth) and ticks the teeth row, so an under-specified tooth is one tap from resolved instead of a dead end. Everything the sheet renders comes from that folded result, not the raw one.
• RTL-safe: logical text-start / text-end only, never text-left/text-right. Dates via lib/i18n/format.ts.

---

8. Gating and configuration

Gate = TAB_TREATMENT_EDIT ∧ canEditTreatmentForDay ∧ configured locale profile ∧ Plan.features.voiceTreatmentEntry. How each failing condition renders is in §2.

Plan.features is a Json column that already exists on the Plan model and is seeded as {} for all five plans — and is read nowhere in backend/src/. Voice is its first consumer, so:

• No migration is needed.
• The read should go through a small generic helper, since this establishes the pattern for every future flag.
• Voice is metered vendor spend, which is why the gate is commercial (plan) rather than a new TAB_* permission — a clinician who can already edit the form gains no capability from voice, only speed.

The per-user throttle is a separate, non-commercial abuse control.

v1 ships ungated

The plan flag is designed but not enforced in v1 — voice is open to every clinic user who can edit treatments, in every configured locale. Plan.features.voiceTreatmentEntry and the availability API stay documented here as the intended gate, deferred rather than dropped, so turning them on later is additive.

Consequence to accept deliberately: with no plan gate and no duration cap (§2), the per-user throttle is the only control on metered vendor spend. See open item 14.

---

9. Errors

Per the three-layer contract: a code in common/errors/error-codes.ts, the throw site, and an errors.X key in all three of frontend/messages/{en,fa,nl}.json. Never a raw English Nest exception for a user-facing failure.

Code	When
VOICE_MIC_DENIED	browser permission refused — client-side only: needs the errors.X key in all three message files, but no ErrorCode entry and no throw site
VOICE_CLIP_TOO_LONG	over maxMs (server-side re-check), or over vendor limits
VOICE_UNSUPPORTED_FORMAT	recorder produced a container the profile rejects
VOICE_ASR_FAILED	transcription stage failed
VOICE_EXTRACT_FAILED	transcript obtained, structuring failed
VOICE_NOTHING_RECOGNIZED	empty or unusable transcript
VOICE_NOT_AVAILABLE	no profile for locale (v1); plan flag off, once enforced
VOICE_RATE_LIMITED	throttle

Transcript salvage: when ASR succeeded and only extraction failed, the response still carries the transcript and the failure dialog offers "افزودن به یادداشت". That action creates a new detail with only comment set to the transcript — everything else left at newDetail() defaults. The words were captured and paid for; only the structure was lost.

This keeps the feature's one invariant intact: voice never writes into an existing detail. Dictating into an already-filled detail is a separate, later feature with its own voice-to-text control scoped to that field (§1, out of scope).

It also does not bypass the confirmation rule — the dialog shows the transcript, and the dentist taps to accept it. That review matters, because a raw transcript carries ASR errors and may contain the patient's spoken name, and comment is persisted (§10).

---

10. Data handling

• Audio is held in memory for the request only. Never written to disk, never a Prisma row. Note this is deliberately unlike treatment attachments, which do persist to backend/uploads/treatments.
• The transcript goes to the browser for the review sheet and dies with it.
• The comment field persists a cleaned version of what was said — that is legitimate clinical record-keeping and is the only durable trace.
• Telemetry is structured and patient-free: clip duration, which fields resolved, unresolved count, vendor latency, usage.cost, outcome (applied / discarded / failed), locale. Never the transcript, never audio, never a patient identifier. The destination is an open question — this repo has no metrics infrastructure yet (open item 9). Emit it as structured log lines in the interim so the fields exist and can be routed later without changing call sites.

Cancelling aborts the in-flight vendor call via AbortController, rather than letting it settle and discarding the result. Note this reduces spend but does not eliminate it: work already performed upstream may still be billed.

Recordings are clinical descriptions of identifiable patients leaving the server for a third party. Confirm the provider's retention and training policy in writing before enabling this for real clinics.

---

11. Open items

1. Persian ASR accuracy on tooth numbers is unmeasured, and it decides the feature. Everything downstream assumes a usable transcript; no design choice above compensates for a bad one. Run this before writing feature code.

   Why tooth numbers specifically, not general accuracy: a transcript can score well on WER and still be useless here, because the errors land on the digits. چهار (4) / چهارده (14) / چهل (40) differ by one syllable. FDI spoken as یک چهار may return as ۱۴, 14, or یک چهار. Persian and Latin digit scripts mix within one transcript. Jargon is loanwords (زیرکونیا, پرسلن فیوزد تو متال, اینله/آنله) and clinicians code-switch into English mid-sentence. Suction and handpiece run in the background.

   Protocol: ~25 utterances from a dentist reading a script covering explicit FDI, quadrant-relative phrasing, bridges, prosthesis types, due dates and notes — recorded on the real device, ideally once quiet and once with the operatory running. POST each to /api/v1/audio/transcriptions with language: "fa". Score per-tooth-reference accuracy (of every tooth spoken, how many survive recoverably?) and jargon recognition separately. Notes accuracy barely matters. ~6 minutes of audio ≈ $0.04.

   Decision it drives: if tooth accuracy holds, build. If not, the fix is not prompt tuning — it is pointing the fa ASR slot at a Persian-specialist vendor while en/nl keep whisper-1. That is a config change precisely because of decision 23.

   Unverified lead for that fallback: recent Persian-ASR benchmark work reports Qwen3-Omni as the strongest open Persian ASR. Not confirmed as available on OpenRouter's transcription endpoint — check before relying on it.

   Byproduct: the recordings become the fixture corpus for item 2.

2. VOICE_LLM_MODEL — resolved: google/gemini-3.7-flash (§4). Still worth validating on real whisper output during item 1, since clean transcripts flatter every model; the escalation path if Persian comprehension disappoints is documented in §4 and is a config change.

3. Production reachability — resolved: confirmed reachable from the Iranian production host. The per-locale registry is retained regardless (decision 23), so the fa profile can still be repointed at a domestic vendor if item 1 goes badly.

4. Provider retention/training policy — resolved: restricted via OpenRouter's account-level privacy/data-policy settings. Note this is an account setting, not a per-request one: re-verify it if the API key or the OpenRouter account changes, and remember whisper-1 is forwarded to OpenAI, so the effective policy is OpenRouter's plus that provider's.

5. English tooth numbering — resolved: a bare two-digit number is read as FDI in all three locales. FDI is what the product is built on and what clinicians dictate. Known trade-off, accepted: a clinician trained under Universal numbering says "tooth 14" and means a different tooth, so an en clinic needs either training or a later per-org notation setting. Revisit if a US clinic is onboarded.

6. nl and en have no spike data. The Persian spike (item 1) should be repeated per locale before that locale's mic is enabled for real users — same protocol, same scoring, different speaker.

7. Transcript salvage has no target — resolved: salvage creates a new detail with only comment set to the transcript (§9). Preserves the invariant that voice never writes into an existing detail; dictating into a filled detail becomes a separate later feature with its own field-scoped control.

8. Throttle limits — resolved: configurable, v1 default 6 requests / 60s per user (§3). Unreachable by a human; a pure abuse guard.

9. Telemetry has no sink. §10 defines exactly what to record but not where it goes — this repo has no metrics or analytics infrastructure yet. Deliberately deferred until it does. Interim: structured log lines, so the fields exist and can be routed later without touching call sites.

10. Review sheet modality — resolved: modal on desktop, full-screen overlay on mobile (§7). Not a route — navigating would unmount TreatmentWorkspace and destroy the draft.

11. Cancel and in-flight requests — resolved: abort via AbortController to limit spend (§10). Upstream work already performed may still be billed.

12. Recording duration cap — resolved: 2 minutes, configurable via maxMs (§2). Still worth timing a realistic worst-case prosthesis dictation during item 1 to confirm 2 minutes is comfortable rather than tight.

13. Availability API — resolved for v1: voice ships open to everyone with a configured locale profile. No plan check, no availability endpoint. The Plan.features design in §8 is deferred, not dropped.

New

14. Cost exposure is bounded but ungated in v1. With the 2-minute cap (12) and the 6/60s throttle (8), worst case is 12 audio-minutes per user-minute ≈ $0.072/min, or ~$4.30 per hour of sustained abuse by one user — bounded, not free. Normal use is far below this: a clinician doing 60 recordings a day at ~20s each costs about $0.18/day.

    What remains open is that there is no per-organization limit at all, because v1 ships ungated (13). Worth deciding what to watch and at what number to react. Mitigations, already designed and each a config or flag change: enforce the Plan.features gate, lower maxMs or the throttle, or add an org-level monthly minute budget.

---

12. Verification

• cd backend && npm test — new suites for resolveToothIntent (quadrant mapping in all four quadrants, out-of-range rejection, deciduous → unresolved), resolveDueDate (per-locale week start, "this" vs "next" weekday, Jalali leap year, month-end), the Jalali port, prosthesis expansion + completeness, and connected-span validation.
• cd backend && npm run build — cross-cutting backend gate.
• cd frontend && npx tsc --noEmit — frontend gate.
• Manual: fa locale, editable day, prosthesis detail with a bridge, dispatch to a linked lab. Then specifically:
  • past day → both segments disabled, control still split (not absent);
  • locale with no profile → control renders unsplit, identical to today;
  • fa vs en → mic sits at the logical end in both, on the same side as the chip's trash;
  • cancel mid-recording → chip strip unchanged, no orphan detail;
  • confirm → always appends a new chip, whatever the active detail contains;
  • Add half → behaves exactly as it did before this change;
  • tap 🎤 during the Lab wizard step → confirm returns to the treatment step;
  • no layout shift in the header row on record start, stop, or the 2:00 auto-stop;
  • hold past 2:00 → auto-stops and proceeds to processing, not an error;
  • cancel during processing → the vendor request is actually aborted;
  • review sheet on mobile → full-screen overlay; closing it leaves the draft intact.

---

13. Decision log

Settled in a grilling session on 2026-08-20.

#	Question	Decision
1	Scope	Everything including lab dispatch
2	AI supply chain	Domestic provider originally; OpenRouter for v1, registry keeps both open
3	Apply model	Review sheet, then apply
4	Speech → FDI	LLM emits intent, code resolves. A spoken number is the FDI code (revised 2026-08-21, §6)
5	Cardinality	One detail per recording
6	Lab destination	Closed list of linked labs, explicit confirm, unticked when inexact
7	Due date	Intent + deterministic resolver
8	Resolver location	Backend, Jalali math ported
9	Prosthesis	Default type + overrides, all-or-nothing
10	Retention	Discard audio and transcript, non-PHI telemetry only
11	Capture	Tap to start/stop, hard cap (see 27)
12	Failure UX	Stage-aware codes, transcript salvage
13	Locales	Provider registry per locale; all three locales enabled
14	Reachability	Registry now, slots filled per deployment
23	ASR model	openai/whisper-1 for every locale; registry kept so fa can diverge
24	Extraction model	google/gemini-3.7-flash; escalation path documented in §4
25	Salvage target	Creates a new detail with only comment set — voice never writes into an existing detail
26	Throttle	Configurable; v1 default 6 requests / 60s per user
27	Duration cap	2 minutes, configurable via maxMs
28	Review sheet	Modal on desktop, full-screen overlay (not a route) on mobile; candidate chips are its only interactive part
29	Cancel	Aborts the in-flight vendor call
30	v1 gating	Open to everyone; Plan.features gate deferred, not dropped
15	Gating	Plan.features flag — its first consumer

UI placement settled in a second grilling session on 2026-08-20.

#	Question	Decision
16	Mic action	Independent record action; the Add half's logic is untouched
17	Record UI	Mic segment toggles ▶/■; inline bar below the header row
18	Unavailable	Absent for technical + commercial; disabled for contextual
19	Component API	One optional voice object prop; undefined means absent
20	Side	Logical end (border-s), exactly like the chip's trash
21	Creation	On confirm, never on tap
22	Creation rule	Confirm always appends a new detail — no blank-reuse guard