Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS
Music renderer benchmark evidence
Performance and memory changes to local music renderers need evidence that the full result still works. A short clip, duration number, or aggregate spectral metric can miss audible chunk-boundary failures, so technical checks and a human listen are separate gates.
What a benchmark records
scripts/music_benchmark.py writes a small JSON report containing:
- prompt and lyrics shape only (character, word, and line counts); the text is never copied into the report;
- fixed seed, requested duration, renderer/backend version, profile name, elapsed time, and peak VRAM;
- WAV format, actual duration, RMS/peak/clipping/silence measures, a broad spectral profile, and a SHA-256 output checksum;
- technical validation status plus an explicit
listeningRequired: truereview marker.
Generated audio and user prompts stay outside Git. Reports are safe to attach to a local investigation because they contain no audio bytes or prompt text.
Technical check
Run the analyzer after a fixed-seed render. Replace the example values with the profile and environment being tested:
python scripts/music_benchmark.py \
--audio /path/to/example.wav \
--output /path/to/example-report.json \
--prompt-file /path/to/example-prompt.txt \
--lyrics-file /path/to/example-lyrics.txt \
--duration 180 \
--backend-version diffusers-example \
--profile balanced \
--elapsed-ms 123456 \
--peak-vram-mb 24576 \
--seed 17
The analyzer expects 16-bit PCM WAV and the shared 32 kHz library rate. It rejects malformed or truncated files, unexpected duration, near-silence, excessive clipping, and catastrophic broad-band drift against an optional baseline report. A rejected report is a failed technical gate. A passed report does not approve the music.
For CUDA Diffusers MiniMax Music 3, --seed is a sidecar-only benchmark hook:
python scripts/generate_minimax_music3.py \
--model MiniMaxAI/MiniMax-Music3 \
--text "Example instrumental prompt" \
--lyrics "[intro]\n[instrumental]\n[outro]" \
--duration 180 \
--output /path/to/example.wav \
--seed 17
The sidecar passes the seed only when the installed Diffusers pipeline exposes a
generator argument. Ordinary PortOS generation does not pass a seed and is
unchanged. If the installed pipeline has no deterministic generator support,
the seeded benchmark exits clearly instead of claiming repeatability.
Required full-length listening review
After technical checks pass, listen to the entire output from beginning to end. Repeat the comparison with the prior supported profile using the same prompt shape, lyrics shape, duration, and fixed seed. Record:
- whether the song reaches the intended ending without a cut or repeated chunk;
- whether transitions, rhythm, stereo image, and loudness remain coherent;
- whether the new profile introduces clicks, silence, clipping, stitching seams, or audible spectral changes;
- the reviewer, date, profile, seed, and a short outcome in the local report or release evidence.
Only a report with passing technical checks and an explicit positive full-length listening result may support marking a renderer profile as supported. Numerical checks never substitute for that listen.