@gotcos/glasses-server 6.61.0 → 6.61.1

package/CHANGELOG.md

@@ -1,3 +1,16 @@
+## 6.61.1
+
+### Fixes 6.61.0 labelling other people as the wearer: a split track is named only by a confident voiceprint, never by the glasses' default label
+
+- Miles, 2026-10-02 10:19, in a live meeting on 6.61.0: "MU is being assigned to loudest speaker vs my actual voice signature ... I saw Jeremy Sokolic voice showing up as MU." In that session (`meeting_1790953219564_yhwyxs`, read from a copy), 40 turns read MU. 25 were genuine (voiceprint at 0.65 or more, Even Hub `self`). 11 had similarity 0 in two-track chunks where Even Hub heard mostly someone else ("Okay.", "I can send a message. Yeah.", "to have as we play into calendar."): a short second track took the client's label, and the client's label is the glasses' default wearer label (`MU`), not evidence of who spoke. 2 more ("Of course. All right." at 0.612, and 0.625) were a track's own voiceprint check returning the wearer from the full search, which accepts any profile from 0.55, below the 0.65 at which the voiceprint verifies the wearer.
+- **The client's label never names a track.** In a split chunk, a short track that is not the dominant one is `Ext`. The client's label is still what 6.60.0 used it for: a whole chunk no voiceprint could run on (under 2 s, or no voiceprint model). A one-track chunk is labelled exactly as 6.60.0 labelled the chunk.
+- **The wearer needs owner-verify confidence in a split chunk.** A track named the wearer (`owner_speaker_label`) by its own voiceprint check, or a short dominant track that would take the whole chunk's voiceprint, keeps the wearer's name only at `OWNER_VERIFY_SIMILARITY` (0.65, now one constant shared with the voiceprint's owner verify); between 0.55 and 0.65 it is `Ext`. Any other name keeps the search threshold, as before. The two 0.665 and 0.716 turns where Even Hub heard mostly someone else are above the floor and stay, as reviewed.
+- **The final pass applies the same rule.** Its channel map was already built only from chunks at 0.65 or more. Its per-chunk fallback (the live turn a word sat in, else the chunk's label) now treats a wearer name in a split chunk below 0.65 as `Ext`, so a 6.61.0 turn like these is never carried into a final transcript.
+- **Logged and counted.** Every chunk where a would-be wearer name was refused logs one line, `[diarizer] owner-guard chunk #N refused=K client_label=X weak_owner=Y`; the chunk's `diarizer.ownerGuard` lists the reasons, and `readiness.diarizer.ownerGuard` counts them since start.
+- **Replayed on the field session (read-only copies of its 195 Nemotron chunks and their audio, the real CLI and the real voiceprint):** all 28 genuine wearer turns (0.65 or more, Even Hub `self`) stay MU; 0 of the 11 similarity-0 MU turns and 0 of the 2 weak ones (0.612, 0.625) are MU any more; the 2 borderline turns stay MU; the guard refused 13 client labels and 2 weak wearer matches; no chunk fell back.
+- Tests: the field chunks rebuilt from their records ("Okay." with Even Hub self 19/121, "I can send a message. Yeah." at self 3/125 with Kyle keeping his own track, "Of course. All right." at 0.612), the wearer staying the wearer in a one-track chunk (0.703) and a split one (0.716), a weak wearer match refused on both paths (the whole-chunk voiceprint a short dominant turn takes, and a track's own check), the client label never naming a track, the owner-guard log line and count, the route passing `owner_speaker_label`, and the final pass refusing a 6.61.0 similarity-0 MU turn while keeping a confident one. The mutation gate adds a case for each guard and retargets the three 6.61.0 cases the guard replaced: baseline 169 tests in 11 files, 104 of 104 killed. Full suite 5,723 passed, 2 skipped.
+- No Install hooks and no COS Control change.
+
 ## 6.61.0
 
 ### Nemotron is the default diarizer, for real: it splits two-voice chunks on the HUD during the call and in the final transcript, and the voiceprint names each voice

package/README.md

@@ -597,8 +597,9 @@ is split into turns instead of carrying one label.
   `segments: [{ speaker, text, startSec, endSec, similarity }]` in time order;
   their texts joined with one space equal `text`. A track with at least 2 s of
   its own audio is named by the voiceprint on that audio; a shorter one takes the
-  chunk's voiceprint when it is the dominant voice, otherwise the client's label
-  (when that is a real name different from the dominant one) or `Ext`. The
+  chunk's voiceprint when it is the dominant voice, otherwise it is `Ext`. The
+  glasses' default wearer label never names a track, and in a split chunk the
+  wearer is named only at owner-verify confidence (0.65; since 6.61.1). The
   segments are stored on the chunk, so the saved meeting keeps them.
 - **The final transcript.** When a meeting is saved, one Nemotron pass over the
   whole meeting maps each channel to a name from the chunks the voiceprint was

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gotcos/glasses-server",
-  "version": "6.61.0",
+  "version": "6.61.1",
   "description": "COS Glasses: self-hosted AI heads-up-display server for Even G2 smart glasses, powered by Claude Code, Codex, Cursor Agent CLI, or local Ollama",
   "type": "module",
   "bin": {

package/server/lib/nemotron-final.ts

@@ -41,6 +41,8 @@ import {
   prepareNemotronModelsDir,
 } from './diarizer-backend.js'
 import { CONFIDENT_SIMILARITY, isUnattributed } from './meeting-speaker-review.js'
+import { getOwnerSpeakerLabel } from './profile.js'
+import { OWNER_VERIFY_SIMILARITY } from './voiceprint-floor.js'
 import {
   NEMOTRON_TMP_PREFIX,
   NemotronRunError,
@@ -265,6 +267,10 @@ export function crowdIdentities(windows: TimelineWindow[]): number {
   return [...counts.values()].filter(count => count >= FINAL_MIN_SUPPORT).length
 }
 
+function ownerLabelOrNull(): string | null {
+  try { return getOwnerSpeakerLabel() } catch { return null }
+}
+
 function isName(label: unknown): label is string {
   return typeof label === 'string' && !isUnattributed(label.trim())
 }
@@ -332,7 +338,21 @@ export function mapChannelsToNames(windows: TimelineWindow[], a: Activity, optio
 }
 
 /** The live segment that held chunk-relative time t, else null. */
-function liveSpeakerAt(chunk: SidecarChunk, t: number): string | null {
+/**
+ * 6.61.1 owner guard for per-chunk names. 6.61.0 named split tracks as the wearer
+ * from the glasses' default label (similarity 0) or a weak search match
+ * (0.55 to 0.65). In a chunk with more than one live turn, the wearer's name
+ * counts only at owner-verify confidence; otherwise it is Ext. A one-turn chunk
+ * keeps its label exactly as 6.60.0 gave it.
+ */
+export function guardOwnerName(name: string, similarity: unknown, chunk: SidecarChunk, owner: string | null | undefined): string {
+  if (!owner || name !== owner) return name
+  const split = Array.isArray(chunk.segments) && chunk.segments.length > 1
+  if (!split) return name
+  return Number(similarity) >= OWNER_VERIFY_SIMILARITY ? name : 'Ext'
+}
+
+function liveSpeakerAt(chunk: SidecarChunk, t: number, owner?: string | null): string | null {
   const segments = chunk.segments
   if (!Array.isArray(segments) || segments.length === 0) return null
   let best: ChunkSpeakerSegment | null = null
@@ -344,7 +364,7 @@ function liveSpeakerAt(chunk: SidecarChunk, t: number): string | null {
     const d = t < s ? s - t : t > e ? t - e : 0
     if (d < distance) { distance = d; best = segment }
   }
-  return best && typeof best.speaker === 'string' ? best.speaker : null
+  return best && typeof best.speaker === 'string' ? guardOwnerName(best.speaker, best.similarity, chunk, owner) : null
 }
 
 export interface RelabelStats {
@@ -383,6 +403,7 @@ export function relabelBatchSegments(
   timeline: MeetingTimeline,
   a: Activity,
   map: Map<number, string>,
+  owner?: string | null,
 ): { segments: SidecarBatchSegment[]; stats: RelabelStats } {
   const stats: RelabelStats = { total: 0, relabeled: 0, viaChannel: 0, kept: 0 }
   const out = segments.map(segment => ({
@@ -399,7 +420,8 @@ export function relabelBatchSegments(
       }
       if (name === null) {
         stats.kept++
-        name = (start && liveSpeakerAt(start.window.chunk, start.inChunk)) ?? original
+        name = (start && liveSpeakerAt(start.window.chunk, start.inChunk, owner))
+          ?? (start ? guardOwnerName(original, start.window.chunk.similarity, start.window.chunk, owner) : original)
       }
       if (name !== word.speaker) stats.relabeled++
       const next: SpeakerWord = { ...word, speaker: name }
@@ -441,6 +463,7 @@ export function relabelStreamingChunks(
   timeline: MeetingTimeline,
   a: Activity,
   map: Map<number, string>,
+  owner?: string | null,
 ): { finals: Map<number, { speaker: string; segments: ChunkSpeakerSegment[] }>; stats: RelabelStats } {
   const stats: RelabelStats = { total: 0, relabeled: 0, viaChannel: 0, kept: 0 }
   const finals = new Map<number, { speaker: string; segments: ChunkSpeakerSegment[] }>()
@@ -449,17 +472,17 @@ export function relabelStreamingChunks(
     if (typeof chunk.text !== 'string' || !chunk.text.trim()) continue
     const tokens = splitTokens(chunk.text)
     const times = streamingTokenTimes(chunk, tokens, window, a)
-    const base = typeof chunk.originalSpeaker === 'string' ? chunk.originalSpeaker : chunk.speaker
+    const base = guardOwnerName(typeof chunk.originalSpeaker === 'string' ? chunk.originalSpeaker : chunk.speaker, chunk.similarity, chunk, owner)
     const labels: string[] = []
     let carry: string | null = null
     for (let i = 0; i < tokens.length; i++) {
-      if (!isWordToken(tokens[i]) || !times) { labels.push(carry ?? liveSpeakerAt(chunk, 0) ?? base); continue }
+      if (!isWordToken(tokens[i]) || !times) { labels.push(carry ?? liveSpeakerAt(chunk, 0, owner) ?? base); continue }
       stats.total++
       const span = times[i]
       const channel = channelForSpan(a, window.start + span.start, window.start + Math.max(span.end, span.start + 0.01))
       let name: string
       if (channel !== null && map.has(channel)) { name = map.get(channel)!; stats.viaChannel++ }
-      else { name = liveSpeakerAt(chunk, span.start) ?? base; stats.kept++ }
+      else { name = liveSpeakerAt(chunk, span.start, owner) ?? base; stats.kept++ }
       labels.push(name)
       carry = name
     }
@@ -646,6 +669,8 @@ export interface FinalPassOptions {
   tmpRoot?: string
   log?: (line: string) => void
   now?: () => number
+  /** The wearer's label, for the 6.61.1 owner guard (default: the profile's). */
+  owner?: string | null
   /** Seam: how files are written (durable atomic by default). */
   write?: (path: string, data: string) => void
   /** Seam: the batch relabeller (word-aligned in place by default). */
@@ -758,6 +783,7 @@ async function finalPassNow(options: FinalPassOptions): Promise<FinalPassOutcome
 
   const sidecarText = readFileSync(options.sidecarPath, 'utf8')
   const sidecar = JSON.parse(sidecarText) as Record<string, any>
+  const owner = options.owner !== undefined ? options.owner : ownerLabelOrNull()
   if (Number(sidecar.correctionRevision ?? 0) > 0) return { status: 'skipped', reason: 'human_corrected' }
   const markdown = readFileSync(options.meetingPath, 'utf8')
   const transcript = readTranscriptSection(markdown)
@@ -833,7 +859,7 @@ async function finalPassNow(options: FinalPassOptions): Promise<FinalPassOutcome
     mode = 'batch'
     const segments = (Array.isArray(sidecar.batchSegments) ? sidecar.batchSegments : []) as SidecarBatchSegment[]
     if (!segments.length) return { status: 'skipped', reason: 'no_speaker_words' }
-    const relabeled = relabelBatchSegments(segments, timeline, activity, mapping.map)
+    const relabeled = relabelBatchSegments(segments, timeline, activity, mapping.map, owner)
     stats = relabeled.stats
     // finalizeBatch wrote labelled turns only when every segment had speaker words.
     const selected = selectBatchTranscriptForPersistence(
@@ -850,7 +876,7 @@ async function finalPassNow(options: FinalPassOptions): Promise<FinalPassOutcome
     if (typeof sidecar.batchTranscript === 'string') nextSidecar.batchTranscript = next
   } else {
     mode = 'streaming'
-    const relabeled = relabelStreamingChunks(timeline, activity, mapping.map)
+    const relabeled = relabelStreamingChunks(timeline, activity, mapping.map, owner)
     stats = relabeled.stats
     const textEntries = entries.filter(entry => typeof entry.chunk?.text === 'string' && entry.chunk.text.trim())
     next = relabelStreamingTranscript(transcript, textEntries.map(entry => ({ chunkIndex: entry.chunkIndex, text: entry.chunk.text })), relabeled.finals)

package/server/lib/nemotron-live.ts

@@ -55,6 +55,7 @@ import {
   type TokenTime,
 } from './nemotron-segments.js'
 import type { WhisperWord } from './whisper-local.js'
+import { OWNER_VERIFY_SIMILARITY } from './voiceprint-floor.js'
 
 /** The chunk response contract shared with COS Glasses 6.9.567. Keep this shape. */
 export interface ChunkSpeakerSegment {
@@ -78,8 +79,10 @@ export interface LiveNamingInput {
   audio: Buffer
   /** Today's whole-chunk voiceprint result. */
   chunkVoiceprint: VoiceprintResult
-  /** The client's own label (amplitude / Even Hub), e.g. the owner or Ext. */
+  /** The client's own label (amplitude / Even Hub), e.g. the owner or Ext. Never names a split track. */
   clientSpeaker: string
+  /** The wearer's label (`owner_speaker_label`), for the owner guard. */
+  owner?: string | null
   identify: (wav: Buffer) => VoiceprintResult | null
 }
 
@@ -92,6 +95,8 @@ export type LiveNamingResult =
       tracks: number
       timing: 'words' | 'estimated'
       identified: number
+      /** Would-be wearer names refused in this chunk, by reason. */
+      ownerGuard: { client_label: number; weak_owner: number }
     }
   | { ok: false; reason: 'no_text' | 'no_speech' }
 
@@ -105,14 +110,19 @@ function isRealLabel(label: string): boolean {
 /**
  * Split one chunk's text into named segments. Pure apart from `identify`.
  *
- * Naming rules (2026-10-02 brief):
- *  - one track: it is the whole chunk, so it takes the chunk's voiceprint result;
+ * Naming rules (2026-10-02 brief, 6.61.1 owner guard):
+ *  - one track: it is the whole chunk, so it takes the chunk's voiceprint result,
+ *    exactly as 6.60.0 labelled the chunk (including the client's label when no
+ *    voiceprint could run);
  *  - two or more: each track with at least NAME_TRACK_MIN_SEC of its own audio is
- *    named by the voiceprint on that audio;
- *  - a shorter track takes the chunk's voiceprint when it is the dominant track,
- *    otherwise the client's label when that is a real name different from the
- *    dominant one, otherwise Ext. A name is never invented and the voiceprint's
- *    own thresholds decide every name it gives.
+ *    named by the voiceprint on that audio; a shorter dominant track takes the
+ *    chunk's voiceprint; any other short track is Ext;
+ *  - in a split chunk the WEARER is named only at owner-verify confidence
+ *    (OWNER_VERIFY_SIMILARITY). A full search can return the wearer from 0.55,
+ *    and the client's label is the glasses' default wearer label, not identity
+ *    evidence: on 2026-10-02 both named other people (Jeremy, Kyle) as the
+ *    wearer. Either becomes Ext and is counted (`ownerGuard`).
+ * A name is never invented.
  * The segments' texts joined with one space equal `text` exactly.
  */
 export function nameChunkSegments(input: LiveNamingInput): LiveNamingResult {
@@ -139,6 +149,7 @@ export function nameChunkSegments(input: LiveNamingInput): LiveNamingResult {
   const used = new Set(labels)
 
   const names = new Map<number, VoiceprintResult>()
+  const ownerGuard = { client_label: 0, weak_owner: 0 }
   let identified = 0
   const pcm = readPcm16Mono(input.audio)
   const ordered = tracks.filter(track => used.has(track.channel))
@@ -146,16 +157,25 @@ export function nameChunkSegments(input: LiveNamingInput): LiveNamingResult {
   if (ordered.length === 1) {
     names.set(ordered[0].channel, input.chunkVoiceprint)
   } else {
+    const owner = input.owner?.trim() || null
+    const ownerSafe = (named: VoiceprintResult): VoiceprintResult => {
+      if (!owner || named.speaker !== owner || named.similarity >= OWNER_VERIFY_SIMILARITY) return named
+      ownerGuard[named.similarity > 0 ? 'weak_owner' : 'client_label']++
+      return { speaker: 'Ext', similarity: 0 }
+    }
     for (const track of ordered) {
       if (track.exclusiveSec >= NAME_TRACK_MIN_SEC && pcm) {
         const wav = exclusiveTrackWav(activity, track.channel, pcm)
         const named = wav ? input.identify(wav) : null
-        if (named) { names.set(track.channel, named); identified++; continue }
+        if (named) { names.set(track.channel, ownerSafe(named)); identified++; continue }
       }
-      if (track.channel === dominant) { names.set(track.channel, input.chunkVoiceprint); continue }
+      if (track.channel === dominant) { names.set(track.channel, ownerSafe(input.chunkVoiceprint)); continue }
+      // The client's label is the glasses' default wearer label, not evidence of
+      // who this voice is (6.61.0 named Jeremy Sokolic as MU through it).
       const dominantName = names.get(dominant)?.speaker ?? input.chunkVoiceprint.speaker
       const client = input.clientSpeaker.trim()
-      names.set(track.channel, { speaker: isRealLabel(client) && client !== dominantName ? client : 'Ext', similarity: 0 })
+      if (isRealLabel(client) && client !== dominantName) ownerGuard.client_label++
+      names.set(track.channel, { speaker: 'Ext', similarity: 0 })
     }
   }
 
@@ -196,7 +216,7 @@ export function nameChunkSegments(input: LiveNamingInput): LiveNamingResult {
   let best = -1
   for (const [name, sec] of voiced) if (sec > best) { best = sec; speaker = name }
   const similarity = [...names.values()].find(named => named.speaker === speaker)?.similarity ?? 0
-  return { ok: true, speaker, similarity: round(similarity, 3), segments, tracks: tracks.length, timing, identified }
+  return { ok: true, speaker, similarity: round(similarity, 3), segments, tracks: tracks.length, timing, identified, ownerGuard }
 }
 
 /**
@@ -244,6 +264,8 @@ export interface LiveLabelInput {
   audio: Buffer
   voiceprint: VoiceprintResult
   clientSpeaker: string
+  /** The wearer's label, for the owner guard on split chunks. */
+  owner?: string | null
   identify: (wav: Buffer) => VoiceprintResult | null
   clock?: () => number
 }
@@ -260,10 +282,12 @@ export interface LiveLabels {
     tracks?: number
     timing?: 'words' | 'estimated'
     ms?: number
+    /** 6.61.1: would-be wearer names refused in this chunk (client_label, weak_owner). */
+    ownerGuard?: string[]
   }
   /** What NemotronLiveRuntime.record should log; absent when the voiceprint was chosen by config. */
   record?:
-    | { outcome: 'nemotron'; ms: number; segments: number; tracks: number; timing: string; identified: number }
+    | { outcome: 'nemotron'; ms: number; segments: number; tracks: number; timing: string; identified: number; ownerGuard?: { client_label: number; weak_owner: number } }
     | { outcome: 'voiceprint'; reason: string; ms: number | null; detail?: string }
 }
 
@@ -306,10 +330,15 @@ export async function resolveLiveLabels(input: LiveLabelInput): Promise<LiveLabe
       audio: input.audio,
       chunkVoiceprint: input.voiceprint,
       clientSpeaker: input.clientSpeaker,
+      owner: input.owner,
       identify: input.identify,
     })
     if (!named.ok) return voiceprintOnly(named.reason, ms)
     const changed = named.speaker !== input.voiceprint.speaker
+    const guarded = [
+      ...Array(named.ownerGuard.client_label).fill('client_label'),
+      ...Array(named.ownerGuard.weak_owner).fill('weak_owner'),
+    ] as string[]
     return {
       speaker: named.speaker,
       similarity: named.similarity,
@@ -320,8 +349,12 @@ export async function resolveLiveLabels(input: LiveLabelInput): Promise<LiveLabe
         timing: named.timing,
         ms,
         ...(changed ? { voiceprintSpeaker: input.voiceprint.speaker, voiceprintSimilarity: input.voiceprint.similarity } : {}),
+        ...(guarded.length ? { ownerGuard: guarded } : {}),
+      },
+      record: {
+        outcome: 'nemotron', ms, segments: named.segments.length, tracks: named.tracks, timing: named.timing, identified: named.identified,
+        ...(guarded.length ? { ownerGuard: named.ownerGuard } : {}),
       },
-      record: { outcome: 'nemotron', ms, segments: named.segments.length, tracks: named.tracks, timing: named.timing, identified: named.identified },
     }
   } catch (error) {
     return voiceprintOnly('error', ms, error instanceof Error ? error.message : String(error))
@@ -414,6 +447,8 @@ export interface LiveSnapshot {
   /** Whether the CLI binary is the v0.17.4 build the offline guarantee was verified on. */
   cliKnown: boolean | null
   chunks: { nemotron: number; voiceprint: number; reasons: Record<string, number> }
+  /** 6.61.1: would-be wearer names refused on split chunks since start. */
+  ownerGuard: { client_label: number; weak_owner: number }
   last: { at: string; outcome: 'nemotron' | 'voiceprint'; reason: string | null; ms: number | null } | null
 }
 
@@ -440,6 +475,7 @@ export class NemotronLiveRuntime {
   private modelsDir: string | null = null
   private identity: { path: string; sha256: string; known: boolean } | null = null
   private counts = { nemotron: 0, voiceprint: 0, reasons: {} as Record<string, number> }
+  private ownerGuard = { client_label: 0, weak_owner: 0 }
   private recent: Array<{ outcome: 'nemotron' | 'voiceprint'; reason: string | null }> = []
   private last: LiveSnapshot['last'] = null
   private lastStateLog: string | null = null
@@ -605,7 +641,7 @@ export class NemotronLiveRuntime {
 
   /** Record one chunk's outcome. Run failures are logged per chunk; state reasons once per change. */
   record(chunk: { sessionId: string; chunkIndex: number }, outcome:
-    | { outcome: 'nemotron'; ms: number; segments: number; tracks: number; timing: string; identified: number }
+    | { outcome: 'nemotron'; ms: number; segments: number; tracks: number; timing: string; identified: number; ownerGuard?: { client_label: number; weak_owner: number } }
     | { outcome: 'voiceprint'; reason: string; ms: number | null; detail?: string },
   ): void {
     const at = new Date(this.now()).toISOString()
@@ -615,6 +651,12 @@ export class NemotronLiveRuntime {
       this.pushRecent('nemotron', null)
       this.lastStateLog = null
       this.log(`[diarizer] chunk #${chunk.chunkIndex} nemotron ${outcome.ms} ms tracks=${outcome.tracks} segments=${outcome.segments} timing=${outcome.timing} named=${outcome.identified}`)
+      const guard = outcome.ownerGuard
+      if (guard && guard.client_label + guard.weak_owner > 0) {
+        this.ownerGuard.client_label += guard.client_label
+        this.ownerGuard.weak_owner += guard.weak_owner
+        this.log(`[diarizer] owner-guard chunk #${chunk.chunkIndex} refused=${guard.client_label + guard.weak_owner} client_label=${guard.client_label} weak_owner=${guard.weak_owner}`)
+      }
       return
     }
     this.counts.voiceprint++
@@ -676,6 +718,7 @@ export class NemotronLiveRuntime {
       warm: { state: this.state, ms: this.warmMs, loadSec: this.warmLoadSec, error: this.warmError },
       cliKnown: this.identity ? this.identity.known : null,
       chunks: { nemotron: this.counts.nemotron, voiceprint: this.counts.voiceprint, reasons: { ...this.counts.reasons } },
+      ownerGuard: { ...this.ownerGuard },
       last: this.last,
     }
   }

package/server/lib/speaker-embeddings.ts

@@ -42,6 +42,7 @@ const require = createRequire(import.meta.url)
 const __dirname = fileURLToPath(new URL('.', import.meta.url))
 
 import { DATA_DIR } from './data-dir.js'
+import { OWNER_VERIFY_SIMILARITY } from './voiceprint-floor.js'
 
 export const SPEAKER_MODEL_FILENAME = '3dspeaker_speech_eres2net_sv_en_voxceleb_16k.onnx'
 
@@ -86,7 +87,7 @@ const PROFILES_PATH = resolve(DATA_DIR, 'voice-profiles.json')
 const CALIBRATION_LOG = resolve(DATA_DIR, 'speaker-calibration.jsonl')
 
 // Thresholds
-const VERIFY_THRESHOLD = 0.65
+const VERIFY_THRESHOLD = OWNER_VERIFY_SIMILARITY
 const SEARCH_THRESHOLD = 0.55
 /** The bar `autoEnroll` enrols at. From 6.45.4 also the bar at which a held
  *  group is offered as a person with one click ("high"): the identifier would

package/server/lib/voiceprint-floor.ts

@@ -1,4 +1,12 @@
-// The least audio the voiceprint names a voice from. One source for the
-// whole-chunk check in transcribe-stream and the per-track check in Nemotron
-// naming, so the two can never drift apart.
+// Voiceprint floors shared by every path that names a voice, so they can never
+// drift apart.
+
+/** The least audio the voiceprint names a voice from: the whole-chunk check in
+ *  transcribe-stream and the per-track check in Nemotron naming. */
 export const VOICEPRINT_MIN_AUDIO_SEC = 2.0
+
+/** The similarity at which the voiceprint verifies the WEARER (speaker-embeddings'
+ *  owner verify). A full search accepts any profile from 0.55, so a voice can come
+ *  back as the wearer between 0.55 and 0.65 without ever passing owner verify;
+ *  6.61.1 refuses that for a split track (it named other people as the wearer). */
+export const OWNER_VERIFY_SIMILARITY = 0.65

package/server/routes/transcribe-stream.ts

@@ -9,7 +9,7 @@ import { chmodSync, readFileSync, writeFileSync, existsSync, lstatSync, mkdirSyn
 import { writeFile } from 'node:fs/promises'
 import { resolve } from 'node:path'
 import { fileURLToPath } from 'node:url'
-import { getVocabulary, getOwnerName } from '../lib/profile.js'
+import { getVocabulary, getOwnerName, getOwnerSpeakerLabel } from '../lib/profile.js'
 import { getOpenAIKey } from '../lib/openai-key.js'
 import { getTranscriptionPolicySnapshot, isOpenAIWhisperFallbackReady } from '../lib/transcription-policy.js'
 import { STRANDED_STALE_MS } from '../lib/stranded-sessions.js'
@@ -1908,6 +1908,11 @@ async function transcribeWithServerWhisper(audioBuffer: Buffer, whisperAudio: Bu
   }
 }
 
+/** The wearer's label for the 6.61.1 owner guard; null when the profile cannot say. */
+function ownerLabelOrNull(): string | null {
+  try { return getOwnerSpeakerLabel() } catch { return null }
+}
+
 function identifyChunkSpeaker(audioBuffer: Buffer, sessionId: string, chunkIndex: number, clientSpeaker: string): { speaker: string; similarity: number } {
   const expectedSpeakers = undefined
   const audioDurationSec = Math.max(0, (audioBuffer.length - 44)) / 32000
@@ -2155,6 +2160,7 @@ async function processStreamChunk(opts: {
         audio: audioBuffer,
         voiceprint,
         clientSpeaker,
+        owner: ownerLabelOrNull(),
         identify: wav => {
           const named = identifySpeaker(wav, undefined, { event: 'nemotron-track' })
           return named ? { speaker: named.speaker, similarity: named.similarity } : null