Task 099: Enceladus Speaker Correction + Transcription Data Layers

Status: DONE

Motivation

Human directives:

• エンケラドスにいるのはミューズではなかったと思う
• 話者判定やそれを前提にした分析はもう一度見直すべき
• whisper のどのモデルを使用したかも記載が無い
• 生の文字起こしデータと、話者分離処理のような追加の前処理をしたデータと、全体の文脈から各種修正を加えたデータを区別して表示するべき

Scope

1. Speaker correction: Verify who is on Enceladus in EP05. If not ミューズ, correct the dialogue attribution and any report references
2. Whisper model documentation: Add which Whisper model (size) was used for each episode's transcription
3. Transcription data layers: Implement 3-tier display on transcription pages:

- Layer 1: Raw transcription data (YouTube VTT or Whisper output, unmodified)

- Layer 2: Preprocessed data (speaker diarization, timestamp alignment)

- Layer 3: Context-corrected data (speaker attribution, text corrections)

1. Re-review: Review speaker attribution across all episodes given this correction

Completed Work

Speaker Correction (prior session)

• DONE: ミューズ → エンケラドスの管理人 (EP02 consistent naming)

- Speaker ID: muse → enceladus-keeper

- Updated: ep05_dialogue.json, ep05_speakers.json, ep05.json, ep02.json, infrastructure.json, attitude-control.json

- Raw transcription data (ep05_lines.json, ep03_lines*.json) NOT modified (Whisper output as-is)

Whisper Model Documentation (this session)

• Added whisperModel optional field to EpisodeLines.sourceSubtitle in dialogue-extraction-types.ts
• Added whisperModel optional field to TranscriptionPageData.sourceInfo and additionalSources in report-types.ts
• Updated all 5 Whisper-sourced lines files with "whisperModel": "medium"
• discoverTranscriptions() in build.ts now passes whisperModel through
• Source info card displays model name
• Tab labels show [medium] for Whisper sources
• extract-dialogue-from-whisper.ts accepts --whisper-model flag for future runs

3-Tier Transcription Data Layers (this session)

• Added layer legend card explaining the 3-tier data model on each transcription page
• Tab labels now prefixed with Layer 3: (corrected) or Layer 2: (preprocessed)
• Layer badges with color coding: green (L3), yellow (L2), gray (L1)
• Layer 1 (raw) explained as git-external; Layer 2 is the closest available view
• CSS: .layer-badge, .layer-dl, .meta-note styles
• 15 passing tests including 3 new: model display, layer legend, alt source model