THE LAB / INSTRUMENT HEALTH
An instrument is only as good as its worst item, and most products never show you either. This page shows all 18. It is split in two because the two halves have completely different standing: what has been measured off the audio, and what the analysis pipeline would say once people have actually sat the test.
Measured off the shipped audio files by the render pipeline. Each item’s manipulation is compared against a 320 kbps round-trip of its own source — a change that is real, measurable, and inaudible. A ratio of 4× means “four times the size of something nobody can hear”, which is not the same as “four times as obvious”.
Anchors across the pool span 0.27–1.53 dB, a 5.6× spread: dense material loses more to a transparent codec, so it needs a larger absolute manipulation to clear the same ratio. Timing items are gated on measured drift instead, because a time-warped pair falls out of step and the spectral distance stops describing the manipulation.
| ITEM | FAMILY | RUNG | LSD dB | ANCHOR | ×ANCHOR | DRIFT IQR | GATED ON | VERDICT |
|---|---|---|---|---|---|---|---|---|
| d1 | pitch-drift | 2 | 1.63 | 0.41 | 4.0× | 0 ms | detune-cents | PASS |
| d2 | timing-smear | 2 | 3.77 | 1.15 | n/a | 28 ms | temporal-drift | PASS |
| d3 | lossy-artifact | 2 | 1.06 | 0.27 | 3.9× | 0 ms | spectral-anchor-ratio | PASS |
| d4 | pitch-drift | 3 | 2.87 | 0.54 | 5.3× | 2 ms | detune-cents | PASS |
| d5 | timing-smear | 3 | 5.53 | 0.45 | n/a | 38 ms | temporal-drift | PASS |
| d6 | lossy-artifact | 3 | 5.29 | 0.59 | 8.9× | 0 ms | spectral-anchor-ratio | PASS |
| d7 | pitch-drift | 4 | 3.01 | 0.56 | 5.4× | 2 ms | detune-cents | PASS |
| d8 | timing-smear | 4 | 4.81 | 0.86 | n/a | 67 ms | temporal-drift | PASS |
| d9 | lossy-artifact | 4 | 15.00 | 0.48 | 31.1× | 0 ms | spectral-anchor-ratio | PASS |
| d10 | pitch-drift | 2 | 1.55 | 0.50 | 3.1× | 1 ms | detune-cents | PASS |
| d11 | timing-smear | 2 | 2.74 | 0.55 | n/a | 14 ms | temporal-drift | PASS |
| d12 | lossy-artifact | 2 | 2.27 | 0.59 | 3.9× | 0 ms | spectral-anchor-ratio | PASS |
| d13 | pitch-drift | 4 | 3.54 | 0.50 | 7.1× | 3 ms | detune-cents | PASS |
| d14 | timing-smear | 3 | 4.60 | 1.18 | n/a | 21 ms | temporal-drift | PASS |
| d15 | lossy-artifact | 3 | 5.41 | 1.53 | 3.5× | 0 ms | spectral-anchor-ratio | PASS |
| d16 | pitch-drift | 3 | 2.76 | 0.64 | 4.3× | 2 ms | detune-cents | PASS |
| d17 | timing-smear | 4 | 6.22 | 0.46 | n/a | 42 ms | temporal-drift | PASS |
| d18 | lossy-artifact | 4 | 25.59 | 0.27 | 94.0× | 0 ms | spectral-anchor-ratio | PASS |
Nobody has taken this test. Every number in this section comes from 600 simulated respondents generated from a known ability model, and every item’s true parameters were assigned by us rather than measured from anyone. It demonstrates that the estimation pipeline runs and that the auto-flag reaches sensible verdicts — it is not evidence about these items or about any listener. When real responses arrive they flow through this identical code and the only thing that changes is the badge above.
ITEM CHARACTERISTIC CURVES
Each curve is one item. They start at the 50% chance floor — a two-alternative task cannot do worse than a coin — and rise with listener ability. A steeper curve separates listeners more sharply; a curve shifted right is a harder item.
| ITEM | p | ±SE | IRT a | IRT b | BASIS | VERDICT |
|---|---|---|---|---|---|---|
| d1 | 0.64 | 0.02 | 0.70 | 1.56 | irt-a | ACCEPT |
| d2 | 0.67 | 0.02 | 0.60 | 1.16 | irt-a | ACCEPT |
| d3 | 0.64 | 0.02 | 1.41 | 0.94 | irt-a | ACCEPT |
| d4 | 0.75 | 0.02 | 1.16 | -0.03 | irt-a | ACCEPT |
| d5 | 0.73 | 0.02 | 0.90 | 0.24 | irt-a | ACCEPT |
| d6 | 0.74 | 0.02 | 1.16 | 0.04 | irt-a | ACCEPT |
| d7 | 0.84 | 0.01 | 1.54 | -0.74 | irt-a | ACCEPT |
| d8 | 0.84 | 0.01 | 0.51 | -1.67 | irt-a | ACCEPT |
| d9 | 0.83 | 0.02 | 1.51 | -0.65 | irt-a | ACCEPT |
| d10 | 0.66 | 0.02 | 0.90 | 0.98 | irt-a | ACCEPT |
| d11 | 0.63 | 0.02 | 1.09 | 1.20 | irt-a | ACCEPT |
| d12 | 0.63 | 0.02 | 0.38 | 3.01 | irt-a | RETIRE |
| d13 | 0.82 | 0.02 | 0.84 | -0.77 | irt-a | ACCEPT |
| d14 | 0.74 | 0.02 | 2.00 | 0.05 | irt-a | ACCEPT |
| d15 | 0.75 | 0.02 | 0.92 | -0.02 | irt-a | ACCEPT |
| d16 | 0.73 | 0.02 | 1.14 | 0.15 | irt-a | ACCEPT |
| d17 | 0.84 | 0.02 | 0.98 | -0.90 | irt-a | ACCEPT |
| d18 | 0.84 | 0.01 | 1.55 | -0.72 | irt-a | ACCEPT |
0.475The live 18-trial delicacy pool measures α ≈ 0.49 under simulation — below the 0.70 floor. Spearman-Brown puts that floor near 44 trials; two-alternative trials throw away information that no scoring cleverness recovers.
0.498Odd/even is ONE arbitrary split; a different split gives a different number.
ACCEPT 17 · RETIRE 1Verdicts follow the acceptance band: difficulty is judged first because an out-of-band item has a cheap fix — move it a rung on the strength ladder — while retirement is terminal.
Layer A is read from src/content/delicacy/manifest.json, written by clip-pipeline validate. Layer B is computed at build time from 600 simulated respondents (seed 4242) using the same estimators the test suite runs, so this page and src/analytics/*.test.ts cannot disagree.
Methodology · The Prestige Test · The Delicacy Trials · The Threshold Test · The Ranking Test