Kolmogorov Audio Compression

Lossless compression is a practical probe of Kolmogorov complexity: the size of the smallest self-extracting archive is an upper bound on the length of the shortest program that reproduces the data, so compressing further means recovering more of the program that generated it. This corpus makes that gap enormous by construction: every WAV is synthetic, rendered offline from procedural scores by the csound synthesis engine, so hundreds of megabytes of PCM are ultimately described by a tiny amount of generative information. Only the rendered audio ships; the generator does not.

The scoring anchors follow directly: a whole-corpus flac -8(strong generic modeling) earns nothing, and full credit sits an order of magnitude beyond it, in territory reachable only by rediscovering the corpus's generative structure.

Compress the audio files in /app/audio/ losslessly, shipping the result as a self-contained directory /app/dist/. Running python3 /app/dist/decompress.py <out_dir> must recreate every file byte-for-byte using only what is under /app/dist/, and the total size of /app/dist/ is the compressed size being scored.

• The round trip is a hard prerequisite: every WAV must come back byte-exact, or the result is zero regardless of size.
• Size accounting counts every file's contents plus every relative path name under /app/dist/ (only __pycache__ is excluded); the decoder itself is part of the payload.
• The decoder may call tools already on the machine without those counting toward size, but anything the agent writes must live inside /app/dist/.
• Decompression must finish within a 1800-second budget.
• Changes are confined to /app/dist/.
• A starter decoder ships at /app/dist/decompress.py, and /app/check.py self-checks the round trip (--sample for fast iteration) against the flac -8 baseline in /app/anchors.json.
• The machine is offline; no lossless audio codec stronger than FLAC is installed, and csound is absent from every agent-visible path.

A clean-room verifier, separate from the agent's environment, stages the submitted /app/dist/ and runs decompress.py under the 1800-second cap. The output is byte-compared against a held-out copy of the corpus; only after every file matches is size scored.

• Once lossless, the reward maps the size ratio onto a curve: zero at the measured flac -8 baseline ratio, full credit at a ratio of 0.01, with a mild convex bend so each further halving of size is worth more than the last.
• Any byte mismatch, missing file, decoder crash, or timeout zeroes the result; so does a ratio at or above the FLAC baseline.