project creatortrack scripts voice profile build 20260715T194211

Building a Scripts app (voice-profile-driven script generator + chat tool) for CreatorTrack on branch feat/social-data-layer-and-scripts (worktree forge-suite-wt/warm-slot-1, dev :3067, DB ct_warm_slot_1), scoped to the Gus the Bass brand workspace (ws 85). As of 2026-07-15: Task 16 (Scripts UI, including a TOCTOU race-condition fix in the 'stats-only' regenerate path) and Task 17 (chat generate_script tool) are done and verified live. Task 18 (formal eval: 20 held-out videos, blind judge, profiled-vs-baseline) and Task 19 (docs) remain, followed by two whole-branch reviews before ship.

Key methodology finding: YouTube auto-captions transcribe the entire audio track, not just the creator's speech, so overlaid memes/music/sound-effects get captured as if spoken (one line looked like model hallucination but was actually a real meme soundbite Gus overlaid on a clip). The corpus was quarantined: 18 of 515 videos (17 music-montage + 1 meme-audio Short) marked 'no usable captions,' leaving 497 verified spoken-word transcripts feeding the voice profile.

Why: Justin caught the bad line by reading the human-readable voice card (that's the design intent of making it human-readable), surfacing real corpus contamination that would have taught the generator to mimic meme audio as Gus's voice.

How to apply: Before trusting or extending the Gus voice profile, confirm Task 18's eval ran against the cleaned 497-transcript corpus, not the pre-cleanup 515. When building similar caption-ingest pipelines for other brands/creators, add a speech-density + music-marker quality gate at ingest time rather than discovering contamination after the fact.

[auto-memory session 60d132de-4659-485c-8044-427af0ccce7f, confidence 0.65, mode staged]