data/ layout

What lives in data/. Everything tracked is gold-standard reference material; everything else is regenerated per run and gitignored.

Layout

data/
├── input_corpus_gold_standard/        — the 12 BBC Radio 4 MP3s + everything we extract from them
│   ├── audio_files/                   — the 12 MP3s (sha256 + URLs in manifest.json)
│   ├── manifest.json                  — corpus MP3 manifest (sha256, source URLs)
│   └── input_extraction/
│       ├── episodes/                  — canonical Episode JSONs (one per MP3, post-merge)
│       └── raw_extractions/           — per-extractor JSONs before merge (10 extractors × 12 eps)
│
│   --- ignored (regenerated per run, never tracked) ---
├── generations/YYYY-MM-DD/<run_id>/   — every concept/episode/audio the system has produced
├── reports/                           — pipeline analytics (calibration, patterns, etc.)
├── contrasts/                         — generated contrast-set JSONs
├── calls.jsonl                        — every API call (cost audit + retry log)
├── critique_log.jsonl                 — every critique run
└── calibration_history.jsonl          — drift-detector state

Documentation files (BEST_CANDIDATES.md, lessons.md, learned_prompt_addendum.md) live under docs/. Code that reads lessons.md / learned_prompt_addendum.md writes them to docs/ directly. The benchmark_seeds.txt regression-suite list lives at docs/benchmark_seeds.txt.

Conventions

Run IDs: YYYY-MM-DDTHH-MM-SS__<8-hex>. Daily buckets in generations/ use the date prefix.
DATA_ROOT: configurable via DOUBLE_ACTS_DATA_ROOT (tests redirect to a tmp dir).
Audio filenames: every user-facing wav embeds <runid8>_iter<NN>_<UTC ts> so files keep their provenance after being moved.
calls.jsonl: append-only, every call tagged with run_id — slice per-run with jq.

Reports (`data/reports/` — gitignored, regenerated)

File	Contents	Generator
`calibration_report.md`	Corpus + contrast scores, separation verdict	`pipeline calibrate`
`patterns_report.md`	Cross-corpus structural patterns	`pipeline patterns`
`benchmark_report.{md,json}`	N=20 benchmark distribution	`benchmark`
`corpus_audio_benchmark.{md,json}`	Audio scores on the 12 corpus MP3s	`benchmark.run_corpus_audio_benchmark`
`cross_model_judge_*.{json,md}`	Gemini + Claude ensemble per-axis comparison	`cross-model-judge`
`report_card_*.md`	Aggregate evaluation across hundreds of categories	`report-card`

data/ layout

Layout

Conventions

Reports (data/reports/ — gitignored, regenerated)

Reports (`data/reports/` — gitignored, regenerated)