None was caught by the numbers looking wrong

Gate 2 reports six measurement errors in its own machinery:

• HNR normalisation
• CPP units and noise floor
• degenerate modal reference
• period CV confounded by intonation
• missing known-good contrast
• repair engine flattering its own metric

Every one of them returned plausible numbers. None was caught by the numbers looking wrong. They were caught by synthetic controls, by a contrast against known-good audio, or by listening.

Four out of six

Four of the six were caught by listening alone. That is the sentence I would put on the site. It says something specific about signal measures on speech: they will produce a number for any input, the number will be in a plausible range, and the only thing that reliably contradicted these six was a person deciding the result sounded wrong.

This is the same finding Weft reached from `uptime` and that the `ewi1963` script header reached from a drifted benchmark box, arrived at from a third direction. The instrument does not announce its own failure.