Sound

Two paths. A supplied track must be analyzed, with beat and onset detection that can be wrong. A synthesized score knows its grid, so beats.json is ground truth, written from the timeline's BPM and meter. Every score here is synthesized in lib/synth.py, with oscillators, filters, a reverb send, a 4× oversampled true-peak limiter, and mastering that measures the result with ffmpeg's EBU R128 meter.

Why the −1 dBTP margin exists. The limiter ceiling is −1.6 dBFS, and the WAV peaked at −1.61. After AAC encoding, the MP4 measured −1.5 dBTP: the codec added overshoot. Always measure the final file.

The phone test. The bass roots sit at 41–65 Hz, which phone speakers can't reproduce. A 150 Hz high-pass dropped the mix 2.9 LU. Muting sources one at a time showed the kick, not the bass, held most of that energy. Fixes:

• harmonics on the bass
• a shorter, higher-ending kick with a click
• a slightly louder pad

The drop became 1.7 LU, and it sounded better on a real phone.

Claude can't hear, but it can see. check.py now writes a spectrogram and a waveform image for Claude to read, plus clipping count, DC offset, crest factor and the small-speaker drop. The final judge is still a human listening on a phone and headphones.

Fact-check of the article

• ❌ downbeats = beats[::4] assumes 4/4 and a correct phase; beat_track doesn't find downbeats. Florios is in 3/4.
• ❌ sin(2π(600 + 900t)t) sweeps at 600 + 1800t Hz. Build the phase with a cumulative sum, as synth.py does.
• ⚠️ peak_pick(delta=0.5) on an unnormalized onset envelope is an arbitrary threshold.
• ⚠️ sfx.mjs hard-clips overlapping sounds, with no limiter or loudness target.

A tautology I removed. "sfx ↔ cues 0.00 ms" compared the cue list against itself. It's now labeled internal consistency. The real sync test measures visual changes in rendered pixels against cue times.