What happened
During the fourth round of PhoonoS compression calibration (23 August 2026), a video podcast was measured over a 40-second window. The perceptual score came back high (XPSNR 40.76): this file has bytes to spare; lower the quality target.
The production probe does not sample that way. It takes three 6-second windows spread across the file, skipping the first 10 %. Re-measured with that sampling — same file, same ladder rung, same CRF — the score was 36.11. Almost 5 dB lower. The file was not over-encoded at all.
Why it matters
The first measurement was not wrong in the sense of a bug; it was a credible number pointing in the wrong direction. Acting on it would have degraded 52 minutes of real content to save bytes that were not spare. Nothing would have failed; listeners would simply have got worse audio.
The rule
Before fixing a calibrated target, reproduce exactly the sampling the code will use — same window count, same durations, same initial skip. A measurement taken differently from production is a different experiment.
This is the sister of Engineering Notes #1: there the hidden variable was a library version; here it was where in the file you look. Both produce plausible numbers, and plausibility is what makes them dangerous.
Applied elsewhere
- The EmpoorioChain performance report records exactly which transactions, in which mix, over which block window, were measured — so that a reader can tell whether a number applies to their case.
- The validator audit reads the live validator set rather than sampling a file that looked like one.
From the PhoonoS media-pipeline work, 2026-08-23.