# MNML SHAKEDOWN CORPUS — provenance

**What this file is, for a reader who arrived here cold.** This is the provenance record for `mnml-shakedown`: a corpus of **Creative-Commons-open minimal techno** pulled from archive.org and frozen on **the training box** as the 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, how the held-out split was drawn, and the hash of every artifact the training run consumes.

**Both ends of the window, in UTC.** The licence-and-metadata survey that chose the items ran **2026-08-30 (UTC)**. The audio intake that downloaded and hashed them froze at **2026-08-31T02:22:53Z** on the laptop. This corpus — transfer to the training box, re-verification, measurement, exclusion, split, and dataset freeze — was built and sealed at **2026-08-31T02:42:30Z**, run id **`20260831T023450Z-mnml0`**.

**How it compares to the survey, in one sentence.** The survey predicted 41 items / 208 tracks / ~23.5 h across the three usable licence tiers and the intake returned exactly that with zero drift and zero licence movement; this freeze then subtracted **8 telephone-rate tracks** the survey could not have seen (it never touched audio) and held out **41** more, leaving **159 training tracks**.

---

## Where the corpus came from

**The survey.** `estate/docs/MINIMAL-TECHNO-INTAKE-2026-08-30.md` (run 2026-08-30 (UTC)) — a metadata-and-licence survey only; no audio was fetched. Its two pool queries, run verbatim against archive.org's scrape API with `fl=identifier,licenseurl,title,creator,year`, paged to exhaustion:

```
POOL A  mediatype:audio AND subject:"minimal techno" AND licenseurl:*publicdomain*
POOL B  collection:netlabels AND subject:"minimal techno"
```

Pool A returned 54 items (all retained); Pool B returned 769, of which 37 carried CC0, CC-BY or CC-BY-SA — matched on the licence **path**, so version and jurisdiction suffixes (`/3.0/`, `/4.0/`, `/3.0/us/`, `/2.5/es/`) count as the same licence, while every NC variant, every ND variant, `nc-sampling+`, and every empty licence field was excluded. The pools overlap by 2 identifiers (`30-hcir-berthaJamesSplit-RWT-012`, `ElectRICHual-Predictable-Mixes`), giving **A ∪ B = 89 unique items**, each individually audited. Ruling **D-20260831-01** then cut the pool to the CC0 + CC-BY + CC-BY-SA tiers only, dropping PD-mark and PD-declaration items, the flagged AI item, the duplicate identifiers, and every mix-only item — leaving the **41 items** in this corpus.

**The intake.** `_work/INTAKE-RECEIPT.md` (frozen 2026-08-31T02:22:53Z on the laptop) is the download receipt: 208 files fetched with the user agent `strata2signal-research/1.0 (private research; contact via strata2signal.com)`, **208/208** matching archive.org's own published size and MD5, zero failures. Each item's `licenseurl` was re-read from its own `https://archive.org/metadata/<id>` record at intake time and matched against its tier's required path — **41 of 41 verified, 0 mismatches**. The verbatim metadata records live in `_work/meta/<identifier>.json`; per-item attribution is in `_work/CREDITS.md`. Those files, not this page, are the licence receipt.

## Transfer to the training box, and the re-verification

The intake tree was rsync'd the laptop → the training box into `_incoming/`, then flattened from `raw/<identifier>/<filename>` to `audio/<identifier>__<filename>`. **208** files were placed (208 by hardlink; hard links keep the flat layout byte-identical to the transferred tree without a second 8 GB copy), with **0** name collisions to resolve.

Every flattened file was then re-hashed **from disk on the training box** and compared to the intake `MANIFEST.sha256` entry for its source path: **208/208 matched, 0 mismatched**, 8,565,840,843 bytes total. `manifest.sha256` at this corpus root carries the same digests against the flat layout and is `sha256sum -c`-able from here. Mapping: `_work/flatten_map.json`; receipt: `_work/verify_report.json`.

## Tier totals — as intaken, and after the exclusions

| Tier | Items | Tracks | Hours | GB | Tracks after exclusion | Held out | Training |
|---|---:|---:|---:|---:|---:|---:|---:|
| CC0 | 9 | 61 | 6.49 | 2.836 | 61 | 10 | 51 |
| BY | 9 | 74 | 8.36 | 1.641 | 74 | 16 | 58 |
| BY-SA | 23 | 73 | 8.69 | 4.088 | 65 | 15 | 50 |
| **TOTAL** | **41** | **208** | **23.54** | **8.566** | **200** | **41** | **159** |

Hours and GB are measured on the files on the training box — `ffprobe` durations and filesystem byte counts, not archive.org's metadata.

## What was measured, and how

Three CPU-only passes ran over `audio/` on the training box. **No pass writes or modifies audio** — the corpus on disk is byte-identical to what the hashes above cover.

1. **Format** — `ffprobe` per file: codec, sample rate, bit depth, channels, duration, bitrate.
2. **Loudness** — `ffmpeg -af loudnorm=print_format=json -f null -` per file: integrated LUFS, LRA, true peak, gating threshold. The filtered stream is discarded; only the analysis JSON is kept. **Measurement only.** Tonight's training path has no loudness stage and follows the proven recipe unchanged; these numbers exist so a future bench can ask whether one is worth adding.
3. **Bandwidth ceiling** — `_work/bandwidth_audit.py`, the estate's standing instrument (lesson 15, owed on every corpus): 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. All three sets of numbers are baked per-file into `manifest.json`.

### Sample rate, measured on the training box

| Sample rate (Hz) | Files (all 208) | Files (200 working) | Share of working |
|---:|---:|---:|---:|
| 48,000 | 30 | 30 | 15.0% |
| 44,100 | 170 | 170 | 85.0% |
| 16,000 | 7 | 0 | 0.0% |
| 8,000 | 1 | 0 | 0.0% |

### Codec and bit depth, measured on the training box

| Codec | Files (all 208) | Bit depth | Files (all 208) |
|---|---:|---|---:|
| mp3 | 99 | 16-bit | 102 |
| pcm_s16le | 64 | n/a (lossy) | 99 |
| flac | 43 | 24-bit | 7 |
| pcm_s24le | 2 | — | — |

Channels: **208** files at 2 ch. The two columns above are two independent censuses of the same 208 files, not a cross-tabulation — a codec row and the bit-depth row beside it are unrelated.

### The spreads across the 200 working tracks

| Measure | Min | p25 | Median | p75 | Max | Mean |
|---|---:|---:|---:|---:|---:|---:|
| Integrated loudness (LUFS) | -31.26 | -14.04 | -11.91 | -10.24 | 0.71 | -12.16 |
| Loudness range (LU) | 0.60 | 3.30 | 5.40 | 8.40 | 19.10 | 6.11 |
| True peak (dBTP) | -5.20 | -0.20 | 0.31 | 1.18 | 5.23 | 0.50 |
| Bandwidth ceiling (kHz) | 1.95 | 15.79 | 19.33 | 20.14 | 24.00 | 17.96 |

Bandwidth verdicts across the 200 working tracks: **97** band edge (expected); **84** natural rolloff (keeps falling); **17** PREMATURE LOSSY CUTOFF; **2** no cutoff detected.

**The dullest working files, shown and KEPT.** These carry the lowest measured spectral ceilings in the training-eligible corpus. **None of them is excluded**, and none should be: lesson 3 forbids thresholding on content bandwidth, and every one of these sits in a 44.1 or 48 kHz container that can carry the whole band. They are listed so the number is visible rather than buried — if the adapter comes out dull, this is the first table to re-read.

| Work key | Tier | Codec | Sample rate (Hz) | Ceiling (kHz) | Verdict | Split |
|---|---|---|---:|---:|---|---|
| `ElectRICHual-Predictable-Mixes__2) ElectRICHual - Predictable (sleep mix).mp3` | CC0 | mp3 | 44,100 | 1.95 | PREMATURE LOSSY CUTOFF | training |
| `30-hcir-berthaJamesSplit-RWT-012__1) 30-hcir - Compl3x.mp3` | CC0 | mp3 | 44,100 | 2.24 | natural rolloff (keeps falling) | held out |
| `hardware-techno__anticipation.wav` | CC0 | pcm_s16le | 48,000 | 3.36 | natural rolloff (keeps falling) | training |
| `STE34__01 - 4 kacts of leicnhelled 8er.mp3` | BY | mp3 | 44,100 | 3.89 | band edge (expected) | training |
| `shinji-wakasa-dawn__Shinji Wakasa - Dawn - 05 Sansui-Ryoku.flac` | CC0 | flac | 44,100 | 4.23 | band edge (expected) | training |
| `shinji-wakasa-dawn__Shinji Wakasa - Dawn - 12 Jyakumaku.flac` | CC0 | flac | 44,100 | 4.95 | natural rolloff (keeps falling) | training |
| `STE34__04 - hibachi.mp3` | BY | mp3 | 44,100 | 5.46 | band edge (expected) | training |
| `MonoKraK197AlbertNegredoLithiumSerendipity__(monoKraK197) AlbertNegredo_John Cade.wav` | BY-SA | pcm_s16le | 44,100 | 6.01 | PREMATURE LOSSY CUTOFF | held out |

## The exclusions — a CONTAINER-RATE call, not a content threshold

**8 tracks**, every track of items **TFN223** and **TFN247**, are excluded from **both training and holdout**. They are the F7 files the intake receipt resolved: honestly-encoded lossless streams — real FLAC, real uncompressed PCM — of telephone-rate source.

| Work key | Item | Codec | Sample rate (Hz) | Nyquist (kHz) | Measured ceiling (kHz) |
|---|---|---|---:|---:|---:|
| `TFN223__01GraffitiMechanism-DaybreaknumbfortablyCumbMix.flac` | TFN223 | flac | 16,000 | 8.00 | 0.92 |
| `TFN223__02GraffitiMechanism-AftermarketWarheadspeaceheadsMixdown.flac` | TFN223 | flac | 16,000 | 8.00 | 0.78 |
| `TFN223__03GraffitiMechanism-AlreadyDeadzombieMixdownVersion.flac` | TFN223 | flac | 16,000 | 8.00 | 3.13 |
| `TFN223__04GraffitiMechanism-AftermathatomicNrgFalloutMix.flac` | TFN223 | flac | 16,000 | 8.00 | 0.38 |
| `TFN247__01GraffitiMechanism-DaybreakdaybrokeMix.wav` | TFN247 | pcm_s16le | 16,000 | 8.00 | 8.00 |
| `TFN247__02GraffitiMechanism-AftermarketWarheadsstickyMixwaydown.wav` | TFN247 | pcm_s16le | 8,000 | 4.00 | 4.00 |
| `TFN247__03GraffitiMechanism-AlreadyDeaduntitledEditionMix3.2.wav` | TFN247 | pcm_s16le | 16,000 | 8.00 | 1.50 |
| `TFN247__04GraffitiMechanism-Aftermathm.i.x.Mix.wav` | TFN247 | pcm_s16le | 16,000 | 8.00 | 7.83 |

**Why this is honest, stated plainly.** This is a **container-rate** exclusion: the file's own sample rate puts Nyquist at 8 kHz (4 kHz for the one 8 kHz file), so the genre's top octave — hi-hats, rides, the transient edge of every percussive hit — is **physically absent from the container**, not merely quiet. Upsampling cannot restore it. That is a different thing from **content-bandwidth thresholding**, which lesson 3 forbids: nothing here is judged on how much energy the *content* happens to carry near its own Nyquist, and no 44.1 kHz file was dropped for sounding dull. The test is one number read off `ffprobe` — the container's rate — and it is applied per **item**, so no EP straddles the line. The exclusion is asserted against the measurement before the freeze proceeds: every track of these two items must be sub-standard-rate and no track outside them may be, or the freeze aborts.

The files stay on disk and stay in `manifest.sha256`. Dropping them costs 2.7% of the corpus by duration and keeps the adapter from learning that minimal techno has no treble. Full record: `_work/exclusions.json`.

## The holdout split — D-20260831-03(c)

The rule, verbatim:

> D-20260831-03(c) — ~20% held out, STRATIFIED BY ITEM (an item's tracks are all-in or all-out; no EP straddles): eligible-for-holdout = items with <=8 tracks (the mega-items hardware-techno, onmp215a, unfound88, shinji-wakasa-dawn, braids-flurish-perish are never held out). Group eligible items by tier (CC0/BY/BY-SA); within each tier sort by identifier case-insensitively; walk taking every 3rd item (index%3==0) until that tier's held-out track count reaches ~20% of the tier's post-exclusion tracks; if short after the walk, a second pass at index%3==1. Deterministic, no seed.

It is deterministic and seedless: the same corpus produces the same split on any box, and an item's tracks are **all-in or all-out**, so no EP is split across the boundary and no held-out track shares a session, a mixdown chain, or a mastering pass with a training track.

| Tier | Items after exclusion | Tracks after exclusion | 20% target | Eligible items (≤8 tracks) | Items held | Tracks held | % of tier |
|---|---:|---:|---:|---:|---:|---:|---:|
| CC0 | 9 | 61 | 12.2 | 6 | 4 | 10 | 16.39% |
| BY | 9 | 74 | 14.8 | 7 | 4 | 16 | 21.62% |
| BY-SA | 21 | 65 | 13.0 | 21 | 4 | 15 | 23.08% |
| **TOTAL** | **39** | **200** | — | — | **12** | **41** | **20.5%** |

The walk, tier by tier, showing each pass:

**CC0** — 61 tracks post-exclusion, target 12.2 held. Ineligible (>8 tracks, never held out): `braids-flurish-perish`, `hardware-techno`, `shinji-wakasa-dawn`.
  - pass `index%3==0`: took `30-hcir-berthaJamesSplit-RWT-012`, `in-novation` → 3 tracks held (still short).
  - pass `index%3==1`: took `archivone-black-saturday`, `knolios_moments` → 10 tracks held (still short).

**BY** — 74 tracks post-exclusion, target 14.8 held. Ineligible (>8 tracks, never held out): `onmp215a`, `unfound88`.
  - pass `index%3==0`: took `kahvi019`, `stroboskop-label013`, `unfound91` → 12 tracks held (still short).
  - pass `index%3==1`: took `shoki005g` → 16 tracks held (target reached).

**BY-SA** — 65 tracks post-exclusion, target 13.0 held. Ineligible (>8 tracks, never held out): none.
  - pass `index%3==0`: took `bump188`, `inoQuo070`, `MonoKraK197AlbertNegredoLithiumSerendipity`, `Monokrak203FloatingMind_AMindIsFloating` → 15 tracks held (target reached).

**Named honestly: CC0 lands under 20%.** Its eligible pool holds only 15 tracks, because three mega-items (`hardware-techno` 24, `shinji-wakasa-dawn` 12, `braids-flurish-perish` 10) carry 46 of the tier's 61 tracks and the rule never holds a mega-item out. Both passes ran to exhaustion and took every eligible item they could; CC0 ends at 10 tracks (16.39%). The corpus-wide figure, 20.5%, is inside the ~20% the ruling asks for, and all three tiers are represented in the holdout.

Items held out:

| Item | Tier | Tracks | Creator | Licence (verbatim) |
|---|---|---:|---|---|
| `30-hcir-berthaJamesSplit-RWT-012` | CC0 | 2 | Richard Wilmer | `http://creativecommons.org/publicdomain/zero/1.0/` |
| `archivone-black-saturday` | CC0 | 1 | ArchivOne | `https://creativecommons.org/publicdomain/zero/1.0/` |
| `bump188` | BY-SA | 3 | Kiyoshi Tomehara | `http://creativecommons.org/licenses/by-sa/3.0/` |
| `in-novation` | CC0 | 1 | Kλпξiðλ | `https://creativecommons.org/publicdomain/zero/1.0/` |
| `inoQuo070` | BY-SA | 7 | inoQuo | `http://creativecommons.org/licenses/by-sa/3.0/` |
| `kahvi019` | BY | 2 | Aaron goldbody vs. Luke the wizard | `http://creativecommons.org/licenses/by/1.0/` |
| `knolios_moments` | CC0 | 6 | knolios | `http://creativecommons.org/publicdomain/zero/1.0/` |
| `MonoKraK197AlbertNegredoLithiumSerendipity` | BY-SA | 2 | Albert Negredo | `http://creativecommons.org/licenses/by-sa/3.0/` |
| `Monokrak203FloatingMind_AMindIsFloating` | BY-SA | 3 | Floating Mind | `http://creativecommons.org/licenses/by-sa/3.0/` |
| `shoki005g` | BY | 4 | Shoki Recordings | `http://creativecommons.org/licenses/by/3.0/` |
| `stroboskop-label013` | BY | 6 | stroboskop | `http://creativecommons.org/licenses/by/3.0/` |
| `unfound91` | BY | 4 | jukka-pekka kervinen | `http://creativecommons.org/licenses/by/4.0/` |

**Held-out files are never preprocessed and never appear in `dataset.json`.** The preprocessor discovers its work from the dataset JSON's `audio_path` entries, not by scanning `audio/`, so the holdout cannot leak in through directory discovery. Per-track work keys for both sides are in `split.json`.

## Captions are PLACEHOLDERS

> PLACEHOLDER, per the Sousa/Bach precedent — NOT authored captions. caption = 'minimal techno, ' + <title cleaned>, where <title> is the archive.org per-file title when the item's metadata record carries one, else the filename stem; cleaned = underscores and dashes replaced by spaces, runs of whitespace collapsed to one, ends trimmed. The trigger word is NOT in the caption text: it is applied by the preprocessor from metadata.custom_tag with tag_position=prepend.

Of the 200 working tracks, **153** take their title from the item's own archive.org per-file `title` field and **47** fall back to the filename stem because the record carries no file title. Nobody listened to these tracks and wrote a description — this follows the Sousa and Bach precedent exactly, and the caption text is not evidence about the audio.

## The freeze — every hash tier

Run id **`20260831T023450Z-mnml0`**, minted once at freeze time and used for the dataset directory, the tensor directory, the log directory, and the adapter output name.

| Tier | Artifact | SHA-256 |
|---|---|---|
| 1 — corpus audio | `manifest.sha256` (208 files, `sha256sum -c`-able from `music/corpus/mnml-shakedown`) | *(per-file; self-hash below)* |
| 1b — corpus manifest self-hash | `manifest.sha256` | `0838d191a919caf52e2e9efbc3ba4426964d42bfca2bfc3a9f171832db4c0bb8` |
| 2 — freeze input | `20260831T023450Z-mnml0/dataset.json` | `d26114c0d3828f1b2aaab947db6e676fc5a617cfbe3def62508c1924b8ae8333` |
| 2 — freeze input | `20260831T023450Z-mnml0/split.json` | `26108872a2cdd7a71cf6a7b9db4cb9b75a002151335171079fe07557e30ab5db` |
| 3 — freeze manifest self-hash | `20260831T023450Z-mnml0/manifest.sha256` | `16321accaac17c1875ff8c4b4d343fb4321b3d5c25ea22ff544308e233293ede` |
| 4 — preprocessed tensors | `tensors/mnml-shakedown/20260831T023450Z-mnml0/manifest.sha256` (self-hash) | `b4c9c59b9828a17191e314bce52d5be880fac88aef233ff65d9f671f2db874b1` |

`dataset.json` carries **159 training samples** and nothing else: `metadata` = `custom_tag: "mnml-shakedown"`, `tag_position: "prepend"`, `genre_ratio: 0`. Durations are `min(int(measured duration), 240)` per file — **142 of 159** sit at the 240 s cap and the rest are their own measured length, from 76 s to 240 s. The file was written, read back from disk, and asserted before this page was generated: `genre_ratio == 0`, every duration in `(0, 240]`, every `audio_path` absolute and present, and no held-out or excluded work key anywhere in it.

## Preprocessing

Run on the training box against the pinned trainer (`music/ACE-Step-1.5` @ `14c0211d`, its own `.venv`), model variant **turbo**, `--max-duration 240`, the card pinned by UUID (`GPU-25bc3288-…`) so no index can address the wrong device. The preprocessor discovers its work from `dataset.json`'s `audio_path` entries — it does **not** scan `audio/` — which is the structural reason the holdout cannot leak in.

- **Processed: 159/159**, 159 `.pt` files written to `music/tensors/mnml-shakedown/20260831T023450Z-mnml0`.
- Cross-checked after the run: the 159 tensor stems are exactly the 159 training work keys — **0** held-out stems, **0** excluded stems, **0** training keys without a tensor.
- `manifest.sha256` in the tensor directory covers every `.pt` and verifies 159/159; `manifest.sha256.sha256` self-hashes it (tier 4 above).

## Full inventory

Every file, with its source URL, verbatim licence URL, tier, creator, measured format, loudness and bandwidth ceiling, is in `manifest.json` (one row per file, 208 rows). The tables below are a summary; that file is the record.

| Item | Tier | Tracks | Split | Creator | Licence (verbatim) |
|---|---|---:|---|---|---|
| `30-hcir-berthaJamesSplit-RWT-012` | CC0 | 2 | held out | Richard Wilmer | `http://creativecommons.org/publicdomain/zero/1.0/` |
| `archivone-black-saturday` | CC0 | 1 | held out | ArchivOne | `https://creativecommons.org/publicdomain/zero/1.0/` |
| `braids-flurish-perish` | CC0 | 10 | training | Braids | `https://creativecommons.org/publicdomain/zero/1.0/` |
| `bump188` | BY-SA | 3 | held out | Kiyoshi Tomehara | `http://creativecommons.org/licenses/by-sa/3.0/` |
| `ElectRICHual-Predictable-Mixes` | CC0 | 2 | training | ElectRICHual | `http://creativecommons.org/publicdomain/zero/1.0/` |
| `hardware-techno` | CC0 | 24 | training | David Murillo Diaz | `https://creativecommons.org/publicdomain/zero/1.0/` |
| `in-novation` | CC0 | 1 | held out | Kλпξiðλ | `https://creativecommons.org/publicdomain/zero/1.0/` |
| `inoquo001` | BY-SA | 4 | training | Chuänchö | `http://creativecommons.org/licenses/by-sa/2.5/` |
| `inoquo018` | BY-SA | 4 | training | Ol | `http://creativecommons.org/licenses/by-sa/2.5/` |
| `inoQuo070` | BY-SA | 7 | held out | inoQuo | `http://creativecommons.org/licenses/by-sa/3.0/` |
| `kahvi019` | BY | 2 | held out | Aaron goldbody vs. Luke the wizard | `http://creativecommons.org/licenses/by/1.0/` |
| `knolios_moments` | CC0 | 6 | held out | knolios | `http://creativecommons.org/publicdomain/zero/1.0/` |
| `monokrak134SinAmigosnignAmigo` | BY-SA | 2 | training | Sin Amigos | `http://creativecommons.org/licenses/by-sa/3.0/ch/` |
| `MonoKraK173YannDetroitVSFloatingMind_Dust` | BY-SA | 2 | training | None | `http://creativecommons.org/licenses/by-sa/3.0/` |
| `MonoKraK197AlbertNegredoLithiumSerendipity` | BY-SA | 2 | held out | Albert Negredo | `http://creativecommons.org/licenses/by-sa/3.0/` |
| `Monokrak199FloatingMind_Schni` | BY-SA | 3 | training | Floating Mind | `http://creativecommons.org/licenses/by-sa/3.0/` |
| `Monokrak200FloatingMind_Birthday_Accelerate` | BY-SA | 3 | training | Floating Mind | `http://creativecommons.org/licenses/by-sa/3.0/` |
| `Monokrak203FloatingMind_AMindIsFloating` | BY-SA | 3 | held out | Floating Mind | `http://creativecommons.org/licenses/by-sa/3.0/` |
| `monoKraK204FloatingMind_DiskoChill` | BY-SA | 3 | training | Floating Mind | `http://creativecommons.org/licenses/by-sa/3.0/` |
| `Monokrak205FloatingMind_VoyagerEP` | BY-SA | 3 | training | Floating Mind | `http://creativecommons.org/licenses/by-sa/3.0/` |
| `MonoKraK208FloatingMind_et_si` | BY-SA | 3 | training | Floating Mind | `http://creativecommons.org/licenses/by-sa/3.0/` |
| `Monokrak209FloatingMind_Ready_For_Flying` | BY-SA | 3 | training | Floating Mind | `http://creativecommons.org/licenses/by-sa/4.0/` |
| `Monokrak214FloatingMind_Hidden_Passion` | BY-SA | 3 | training | Floating Mind | `http://creativecommons.org/licenses/by-sa/4.0/` |
| `Monokrak216FloatingMind_Thru_Lines` | BY-SA | 3 | training | Floating Mind | `http://creativecommons.org/licenses/by-sa/4.0/` |
| `Monokrak217FloatingMind_Spatial_Moments` | BY-SA | 3 | training | Floating Mind | `http://creativecommons.org/licenses/by-sa/4.0/` |
| `monokrak84VariousArtistssmokedMonoVol.7` | BY-SA | 3 | training | Zaid Edghaim, Kimo-S, Alicia Hush, Floating Mind | `http://creativecommons.org/licenses/by-sa/3.0/` |
| `Mr.Dee_D1` | CC0 | 3 | training | Mr.Dee | `http://creativecommons.org/publicdomain/zero/1.0/` |
| `onmp215a` | BY | 24 | training | VA | `http://creativecommons.org/licenses/by/3.0/` |
| `RAR005_Derek_Scott_Tomorrow_EP` | BY-SA | 3 | training | Derek Scott | `http://creativecommons.org/licenses/by-sa/2.5/` |
| `shinji-wakasa-dawn` | CC0 | 12 | training | Shinji Wakasa | `https://creativecommons.org/publicdomain/zero/1.0/` |
| `shoki005g` | BY | 4 | held out | Shoki Recordings | `http://creativecommons.org/licenses/by/3.0/` |
| `STE34` | BY | 5 | training | krai | `http://creativecommons.org/licenses/by/3.0/` |
| `stroboskop-label013` | BY | 6 | held out | stroboskop | `http://creativecommons.org/licenses/by/3.0/` |
| `Stroboskop033` | BY | 6 | training | stroboskop | `http://creativecommons.org/licenses/by/4.0/` |
| `TFN110` | BY-SA | 3 | training | Lik-o | `http://creativecommons.org/licenses/by-sa/3.0/us/` |
| `TFN211` | BY-SA | 2 | training | Graffiti Mechanism | `http://creativecommons.org/licenses/by-sa/3.0/us/` |
| `TFN223` | BY-SA | 4 | **excluded** | Graffiti Mechanism | `http://creativecommons.org/licenses/by-sa/3.0/us/` |
| `TFN247` | BY-SA | 4 | **excluded** | Graffiti Mechanism | `http://creativecommons.org/licenses/by-sa/3.0/us/` |
| `tranz023Holocaos-MetamorfoseComputadorEp_140` | BY | 8 | training | Caue Miranda | `http://creativecommons.org/licenses/by/4.0/` |
| `unfound88` | BY | 15 | training | None | `http://creativecommons.org/licenses/by/4.0/` |
| `unfound91` | BY | 4 | held out | jukka-pekka kervinen | `http://creativecommons.org/licenses/by/4.0/` |

---

*Built on the training box, 2026-08-31T02:42:30Z. Corpus root `music/corpus/mnml-shakedown`; dataset freeze `music/datasets/mnml-shakedown/20260831T023450Z-mnml0`. Sources: the survey `estate/docs/MINIMAL-TECHNO-INTAKE-2026-08-30.md` (2026-08-30 (UTC)) and the intake receipt `_work/INTAKE-RECEIPT.md` (2026-08-31T02:22:53Z). No audio file in this corpus was modified at any stage — every pass reads.*