Publish only what you can measure

The mastering case study needed numbers. Twelve pre-masters, twelve reference masters, and a page whose whole argument is that mastering claims should be measurements. The question that shaped the build was not which numbers — it was whose.

Two sources, one honest

The obvious source was WPAudio Engine’s own engineering reports. They exist, they are thorough, and they live in the private mastering workspace — which is exactly the problem. Publishing them would have meant retyping figures the repository cannot verify, that CI cannot check, and that would silently drift the first time a master is revised. A site that unit-tests its contrast ratios cannot ask readers to trust hand-copied loudness values.

The other source was sitting in public/: the actual listening files. Measure those, with a standard meter anyone can run — ffmpeg’s implementation of the BS.1770 family — and every published number becomes reproducible by any reader with the repository and one command. So that is the rule now, recorded as ADR 0007: published measurements are measured from the published files. The engine’s fuller reports — PLR, crest, stereo correlation, DC offset — stay unpublished until they can be published the same way, and the case study says “not published here yet” instead of approximating them.

The rule had teeth immediately. It decided what the Measure section could contain, produced the provenance note that ships in the album catalogue’s source file, and killed a handful of sentences that were true but unprovable.

The pairing had to be earned too

The catalogue maps each original to its reference master. The tempting implementation reads the two folders and matches by filename. But a filename is a claim, not evidence — so the pairing was verified by duration instead: every OG/Master pair matches to the sample, which is what makes them the same song. One detail justified the paranoia: track 09’s filename contains a double space, faithfully preserved in the catalogue with a comment warning future hands not to “fix” it. A tidier filename would be a broken player.

Level-matching, because ears cheat

The A/B instrument switches between versions at the same playhead. Building it surfaced the last honesty problem: the reference masters measure several LU louder, and louder reliably wins a quick switch regardless of quality. An instrument that lets gain win the comparison would undermine the page it sits on. So the player grew a level-match toggle that attenuates the master by that track’s measured loudness delta — the measurement doing double duty as both published data and playback calibration. The loudness article covers why this bias matters and where the research on it lives.

None of this made the page more impressive. All of it made the page checkable, which is the only kind of impressive this Studio is interested in.