# HOUSE-WIDE CORPUS — provenance
**What this file is, for a reader who arrived here cold.** This is the provenance record for `house-wide`: a corpus of **open-licence house and deep house** extracted from the MTG-Jamendo dataset and frozen as the arm-1 training set for an ACE-Step 1.5 LoRA style adapter. It states where every file came from, the licence that permits its use, what was measured on the real bytes, which tracks were excluded and why, and the hash of every file the training run will consume.
**Both ends of the window, in UTC.** The licence-and-metadata survey that chose these tracks ran **2026-09-05 (UTC)**. This extraction — unpacking, hashing, measuring — froze at **2026-09-06T09:10:50Z**, run id **`20260906T080728Z-house-wide`**.
**How it compares to the survey, in one sentence.** The manifest predicted **416 tracks / 35.10 h**; this extraction placed **416** files with zero drift totalling **35.10 h**, of which **416** matched MTG's own published sha256 byte for byte and none disagreed.
---
## Where the corpus came from
**The source.** MTG-Jamendo, read at repository commit `cafd8e20c265ed84f1e61f1c875327971f43a62f`. Track selection, tags and licences come from that commit's `data/` directory; the audio is the ≥30 s ('raw_30s') 320 kbps MP3 set, unpacked from MTG's own tar archives. The manifest that drove this extraction is `house-wide-intake/manifest.jsonl` — it, not this page, is the selection receipt.
**The rulings that shaped it.** D-20260831-01 (CC0 + BY + BY-SA only; NC and ND never) set the licence rung. D-20260905-88 made `raw_30s.tsv` the tag source of record, because the cleaned `autotagging.tsv` deletes the very tags a fence names. D-20260905-91 set the style fence — IN: house, deep house, tech house; OUT: 29 named genre tags plus a title regex. D-20260905-92 removed the tracks whose authors have re-licensed them off the open rung since 2019. Every excluded track is named — the style fence's in `INTAKE-JAMENDO-HOUSE.md`, the relicensed class below and in `CREDITS.md`.
**Coverage.** The manifest's tracks live in **98** of MTG-Jamendo's 100 audio archives; **98** of those archives were on disk when this ran.
## Licence totals — measured on the files on disk
| Licence | Tracks | Hours | GB | Artists |
|---|---:|---:|---:|---:|
| CC-BY | 138 | 13.20 | 1.926 | 26 |
| CC-BY-SA | 278 | 21.90 | 3.171 | 74 |
| **TOTAL** | **416** | **35.10** | **5.097** | **95** |
Hours and GB are `ffprobe` durations and filesystem byte counts on the extracted files — not the catalogue's numbers.
## What was measured, and how
Three CPU-only passes ran over `audio/`. **No pass writes or modifies audio** — the files on disk are byte-identical to what the hashes below cover.
1. **Format** — `ffprobe` per file: codec, sample rate, channels, duration, bitrate.
2. **Loudness** — `ffmpeg -af loudnorm=print_format=json -f null -` per file: integrated LUFS, LRA, true peak. The filtered stream is discarded; only the analysis JSON is kept. **Measurement only** — the training path has no loudness stage.
3. **Bandwidth ceiling** — the estate's standing instrument `bandwidth_audit.py`, unmodified: a 60 s window from the middle of each file, decoded mono at the file's **own** sample rate so no resample can invent a ceiling, 8192-pt Hann frames at 50 % overlap, and the ceiling read as the highest frequency staying within 60 dB of the peak bin.
**Neither loudness nor bandwidth filters anything.** Lesson 3 forbids thresholding on content bandwidth; these columns exist so the numbers are visible rather than buried.
### Sample rate and codec, measured on disk
| Sample rate (Hz) | Files | | Codec | Files |
|---:|---:|---|---|---:|
| 44,100 | 406 | | mp3 | 416 |
| 48,000 | 10 | | — | — |
The two halves of that table are two independent censuses of the same files, not a cross-tabulation.
### The spreads
| Measure | Min | p25 | Median | p75 | Max | Mean |
|---|---:|---:|---:|---:|---:|---:|
| Integrated loudness (LUFS) | -21.96 | -12.37 | -10.21 | -9.09 | -1.33 | -10.93 |
| Loudness range (LU) | 0.40 | 3.70 | 5.35 | 7.60 | 16.90 | 5.83 |
| True peak (dBTP) | -8.01 | 0.27 | 0.88 | 1.59 | 4.70 | 0.93 |
| Bandwidth ceiling (kHz) | 12.54 | 18.70 | 19.91 | 20.06 | 20.38 | 19.12 |
| Duration (s) | 30.0 | 225.7 | 290.8 | 364.7 | 905.2 | 303.8 |
Bandwidth verdicts across the corpus: **413** band edge (expected); **3** natural rolloff (keeps falling). Every one of these files is a 320 kbps MP3 from Jamendo, so a cutoff below the band edge is the encoder, not the music — and nothing is excluded for it.
## Excluded: `relicensed-since-2019` (D-20260905-92)
**17 tracks are absent from this corpus even though their 2019 licence permits training.** MTG-Jamendo's `audio_licenses.txt` records the licence each track was published under and does not change. Jamendo is a live catalogue, and these authors have since re-licensed their own work **off** the CC0/BY/BY-SA rung. Ruling D-20260905-92: they do not train. They are named here rather than quietly dropped, because a corpus that cannot say what it left out cannot be checked.
**A track left only on positive evidence** — its record was read from Jamendo's API and that record's licence is off the rung. A track the census could not read stays in, under its frozen 2019 licence; an unread row is not evidence of anything.
**Census coverage:** 574 of 574 tracks read (290 present, 284 absent from the live catalogue).
| Track | Artist | Licence in 2019 | Licence today | API record read at (UTC) |
|---|---|---|---|---|
| Riot | #Dance 75# | CC-BY-SA | **CC-BY-NC-ND** | 2026-09-06T00:25:29Z |
| By Frank | Andrea Barone | CC-BY-SA | **CC-BY-NC** | 2026-09-06T00:25:40Z |
| Ladiko | Andrea Barone | CC-BY-SA | **CC-BY-NC** | 2026-09-06T00:26:07Z |
| Prassonissi | Andrea Barone | CC-BY-SA | **CC-BY-NC** | 2026-09-06T00:26:09Z |
| Something in the air | Andrea Barone | CC-BY-SA | **CC-BY-NC** | 2026-09-06T00:26:10Z |
| Spirit of Indians | Andrea Barone | CC-BY-SA | **CC-BY-NC** | 2026-09-06T00:26:11Z |
| Tell me | Andrea Barone | CC-BY-SA | **CC-BY-NC** | 2026-09-06T00:26:13Z |
| Trip | Andrea Barone | CC-BY-SA | **CC-BY-NC** | 2026-09-06T00:26:14Z |
| Welcome into my world | Andrea Barone | CC-BY-SA | **CC-BY-NC** | 2026-09-06T00:25:42Z |
| 1.Around | J.O.R.B.I | CC-BY | **CC-BY-NC-ND** | 2026-09-06T00:36:24Z |
| 2.Dancing of the Night | J.O.R.B.I | CC-BY | **CC-BY-NC-ND** | 2026-09-06T00:36:26Z |
| 3.Another Day | J.O.R.B.I | CC-BY | **CC-BY-NC-ND** | 2026-09-06T00:36:27Z |
| 4.Sounds of music | J.O.R.B.I | CC-BY | **CC-BY-NC-ND** | 2026-09-06T00:36:28Z |
| Missing signal | J.O.R.B.I | CC-BY | **CC-BY-NC-ND** | 2026-09-06T00:36:44Z |
| La Vie | On returning | CC-BY | **CC-BY-NC-ND** | 2026-09-06T00:33:52Z |
| In the air | Starlight | CC-BY-SA | **CC-BY-NC-ND** | 2026-09-06T00:27:35Z |
| Tim Perry- Take a chance | Tim Perry | CC-BY-SA | **CC-BY-NC-SA** | 2026-09-06T00:28:30Z |
## Integrity
`manifest.sha256` at this corpus root carries the digest of every file and is `sha256sum -c`-able from here. **416 of 416** also match MTG-Jamendo's own published per-track digest (`data/download/raw_30s_audio_sha256_tracks.txt` at commit `cafd8e20c265ed84f1e61f1c875327971f43a62f`), which is the end-to-end check that the bytes crossing the network are the bytes MTG packed.
Per-track attribution is in `CREDITS.md` beside this file. That file, and MTG-Jamendo's `audio_licenses.txt` behind it, are the licence receipt — not this page.