@gotcos/glasses-server 6.21.4 → 6.21.5

package/CHANGELOG.md

@@ -1,3 +1,12 @@
+## 6.21.5
+
+- Drop the narrow Large-v3 stock-caption artifact family beginning with
+  “closed captioning provided by” from prompt dictation.
+- When that exact HQ artifact occurs, retain the already-validated Fast
+  transcript for the same audio and report the result as degraded instead of
+  saving either the caption credit or an empty gap. All other HQ and no-speech
+  paths are unchanged.
+
 ## 6.21.4
 
 - Preserve validated phone-photo references on Claude and Codex conversation

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gotcos/glasses-server",
-  "version": "6.21.4",
+  "version": "6.21.5",
   "description": "COS Glasses — self-hosted AI heads-up-display server for Even G2 smart glasses, powered by Claude Code, Codex, or Cursor Agent CLI",
   "type": "module",
   "bin": {

package/server/lib/hallucination-filter.ts

@@ -14,8 +14,11 @@
 import { getNegativeRules, getVocabulary, getOwnerName, getWhisperCorrections } from './profile.js'
 
 // ── Whole-chunk silence hallucinations ─────────────────────────────────────
+const CAPTION_CREDIT_HALLUCINATION = /^\s*(?:closed\s+)?caption(?:ing|s)?\s+(?:is\s+)?provided\s+by\b/i
+
 const KNOWN_HALLUCINATIONS = [
   /^subtitles?\s+by\b/i,
+  CAPTION_CREDIT_HALLUCINATION,
   /\blike\s+and\s+subscribe\b/i,
   /\bthanks?\s+for\s+watching\b/i,
   /\bplease\s+subscribe\b/i,
@@ -27,6 +30,13 @@ const KNOWN_HALLUCINATIONS = [
   /^\s*♪+\s*[^♪\n]{0,40}\s*♪+\s*$/,
 ]
 
+/** Narrow stock-caption detector used to retain a valid Fast draft when an HQ
+ * decode leaks a caption credit. It intentionally requires the phrase to begin
+ * the chunk so normal questions about closed captioning remain untouched. */
+export function isCaptionCreditHallucination(text: string): boolean {
+  return CAPTION_CREDIT_HALLUCINATION.test(text)
+}
+
 // ── Inline "Name:" detector (session-aware in streaming, static in one-shot) ──
 const INLINE_HALLUCINATION_THRESHOLD = 3
 // Seed list of known Whisper "Name:" training artifacts. Empty by default — the

package/server/routes/prompt-drafts.ts

@@ -29,6 +29,7 @@ import {
   stripInlineHallucinations,
   stripPromptDictationArtifacts,
   isFullHallucination,
+  isCaptionCreditHallucination,
   isBrandUrlOnly,
   clearSessionHallucinationState,
   applyNegativeRules,
@@ -250,6 +251,35 @@ async function transcribeChunk(draftId: string, chunkIndex: number, audio: Buffe
       return text
     } catch (err) {
       if (err instanceof NoSpeechDetectedError) {
+        // Large-v3 can rarely replace a clipped but valid phrase with a stock
+        // caption credit. Keep the already-decoded Fast text for this one
+        // recognized failure shape instead of persisting either the artifact
+        // or an empty gap. Other no-speech results retain existing behavior.
+        if (mode === 'hq' && isCaptionCreditHallucination(err.rawText)) {
+          const fastJob = modeQualityJobs.get(modeQualityKey(draftId, chunkIndex, 'fast', hash))
+          if (fastJob) {
+            try { await fastJob } catch { /* normal no-speech path below */ }
+          }
+          const fast = loadPromptDraftMeta(draftId)?.warmTranscripts?.[String(chunkIndex)]
+          if (
+            fast
+            && fast.hash === hash
+            && fast.requestedMode === 'fast'
+            && fast.actualQuality === 'fast'
+            && fast.text.trim()
+          ) {
+            if (purpose === 'final') {
+              await markPromptDraftChunkTranscript(draftId, chunkIndex, {
+                ...fast,
+                requestedMode: 'hq',
+                degraded: true,
+                degradationReason: 'hq_caption_credit_hallucination',
+              }, 'final')
+            }
+            console.warn(`[prompt-draft] dropped HQ caption-credit hallucination ${draftId}/${chunkIndex}; retained Fast text`)
+            return fast.text
+          }
+        }
         await markPromptDraftChunkTranscript(draftId, chunkIndex, {
           text: '', hash, requestedMode: mode, actualQuality: mode, backend: 'no-speech', degraded: false,
         }, purpose)