Methodology

Hume closed his essay with a job description: strong sense, delicate sentiment, improved by practice, perfected by comparison, cleared of prejudice — that is a true judge. Standard of Taste's methodology is that sentence turned into engineering constraints.

1. Performance over self-report. Every instrument is a task where you can be wrong. Questionnaires measure your self-image; tasks measure what you actually did. The prestige gap is computed from what your ratings did under false labels; delicacy from whether you found the planted flaw; good sense from whether your confidence matched your accuracy. Nothing asks you to describe your taste, because that answer was never evidence.

2. The user is their own control. Wherever possible the design is within-subject: your labeled ratings are compared to your blind ratings, your retest to your baseline. This removes the need for an external ground truth about which music is good — the instrument never has to take a side in that argument to measure your movement within it. The Prestige Test additionally carries unlabeled control clips, rated in both passes and labeled in neither: they measure each user's plain second-pass drift (memory, familiarity, regression), and the headline score subtracts the residual that drift would leave in it.

3. Deterministic scoring, in code. Every number is computed by a scoring engine whose rules are fixed and inspectable — same responses, same score, every time. No language model classifies you, no black box guesses. Every sentence you read is a fixed template filled in with numbers the engine has already computed; no language model writes any of it when you visit.

4. No number the data can't back. Until a real calibration cohort exists, results carry a provisional label and no percentile appears anywhere in the product. As sessions accumulate, the psychometrics are standard and open about their assumptions: item-response theory for item difficulty and discrimination, signal-detection analysis for the trials, calibration curves and Brier scores for confidence, reliability checks before any norm is published — always with its N attached.

5. The rulers were not invented here. Every figure above is simulated and the cohort is zero, so this product cannot argue from data about people. What it can show is where its measuring apparatus came from — which was published practice all along, uncredited until now.

EBU R 128 — The broadcast method for measuring perceived loudness, rather than peak level. Every clip is loudness-normalised with a two-pass R 128 measurement to one fixed target before it is ever played, so no clip can seem better simply for arriving louder than the one before it. The target here is chosen for headphone listening rather than for broadcast delivery, and this file states the figure it actually uses rather than any figure the standard recommends.

The transparent-encode anchor — A reference point for how much measurable difference an encode can cost while remaining, by consensus, inaudible. Clip fitness is judged against a 320 kbps MP3 round-trip of the same recording rather than against a fixed number, because the same encode costs more spectral distance on dense material than on sparse material. The anchor is a convention rather than a published standard, and it is a measurement of audio files: nobody has listened to confirm the clips it passes are transparent to any actual ear.

ITU-R BS.1534-3 (MUSHRA) — The international recommendation for subjective listening tests, using a hidden reference and anchors, with each sample rated on a continuous scale from 0 to 100. The same document places BS.1116 over small impairments and itself over intermediate quality. It is the tradition our trials sit in rather than a specification we claim to meet: a forced choice between two samples with a required listen, scored against chance. Our pairs carry near-transparent damage, which by that document's own division is BS.1116's regime rather than MUSHRA's — so the familiar 0-to-100 scale is the wrong one to picture here, and we do not use it.

Two of these are decisions living in this repository, so a test opens the file that implements them and fails the build if the passage has moved. The other one rests on a document no test can open, and carries the date a person opened it instead. Those are not the same strength of claim, and this page will not pretend otherwise.

ITU-R BS.1534-3 (MUSHRA) puts a sample anywhere on a continuous scale from 0 to 100. Pitchfork offers 101 places to put a record. Two very different attempts at the same problem, both landing near a hundred degrees. This one offers eleven, and asks only how many of them you used. That last number is the comparison reading, and it is the only place in this product where a professional's scale appears beside your own. It is a reference point and never a target: agreement with a critic is not scored here, because the Prestige Test measures being moved by a prestigious name. A product that measures how far a famous name moves your ratings cannot also give you credit for agreeing with a famous critic.

The analytics dataset behind this — a different store from the one on your device — is self-generated and boring by design: anonymized response vectors under a random session id, carrying ratings, listen times, item-pool version and the scores computed from them. No accounts, no names, no ad-tech. Your browser keeps only your raw answers, never a score; the two are separate on purpose, and the terms say which is which. It exists so the instruments can be calibrated honestly, and that's the whole job. The criteria these rules serve are in the library — start with freedom from prejudice — or skip the theory and take the Prestige Test.