0007 Which samples we ship

Status: accepted

Context

Real audio (0001) needs recordings, and people publish what they make with claudius. Free samples are plentiful; samples whose licence is clear are not. In the sets live coders use every day:

• many drum machine sets trace back to anonymous uploads with no stated terms;
• TidalCycles’ Dirt-Samples has no licence file, and some of its folders are cut from commercial records;
• some well-made libraries are free to use in music but forbid redistributing the samples themselves (Audio Realism’s TR-909 pack, the jRhodes3d Rhodes);
• share-alike licences can reach into the music made with the samples.

Putting samples in the repository or the release archives would also make claudius a redistributor, and the repository huge.

Decision

claudius ships no samples. Sample packs are installed on demand from their authors’ own hosts, at pinned revisions, by claudius setup or claudius-packs install (crates/claudius-packs). The curated registry records, for each pack, its licence, its credit line, whether credit is required, and a licence_note quoting the licence text and where it was read.

• The default set (14 packs, 162 instruments) contains only packs whose licence lets you make and publish music with them, with at most an attribution: CC0, public domain, the Unlicense, CC BY, and free-for-music licences such as jRhodes’ and Audio Realism’s. Packs that forbid redistribution are acceptable because each machine downloads them from their author.
• Credits are automatic: a song that plays CC BY packs gets a credits.txt naming exactly the packs it used.
• Opt-in packs, whose licence is unknown or carries conditions (tidal-drum-machines, dirt-samples, mridangam under CC BY-SA), install only with --accept-licence and print a warning.
• User maps added with add-map record the licence the map or its repository declares, or unknown with a warning that also appears in the credits of every song that uses them.
• Apple Loops are read where GarageBand or Logic installed them, under Apple’s licence for using them in your own music, and never copied, except the drum one-shots claudius cuts into its cache.
• Where no clearly licensed recording exists, claudius does without or builds one: the choir is a single sung vowel, CC0, played as a five-voice ensemble, and the uzu kit’s processed Amen break is left out.

Rejected: bundling packs (redistribution, and a gigabyte-sized repository); enabling the popular unlicensed sets by default (users would publish samples they have no right to); limiting claudius to CC0 (it would lose the best piano, the electric pianos and the acoustic kit, all CC BY).

Consequences

• Music made with the default set can be published, with the credits claudius writes.
• Every licence decision can be checked by reading registry.json, without repeating the research.
• The first run downloads about 2.2 GB from a dozen hosts, and claudius depends on those hosts staying up. GitHub-hosted packs are pinned by commit; the two archives (the 909 and the FreePats organ) are not pinned by hash, so a compromised host could serve different audio (Security).
• Strudel’s default drum-machine banks, which Strudel users expect, are opt-in.
• The library has gaps where no clearly licensed recording exists: no multi-sampled choir, no vocals with words, a piano cut to 5 of its 16 layers for size.
• Adding a pack means reading its licence at the source and quoting it (Sample packs).