An instrument that cannot abstain is not measuring.

A system that always emits a number misleads whenever noise rivals signal. Language models give unstable answers to legal questions even at deterministic settings, so any instrument built from them has a noise floor, and a reading below that floor is not information about the law. It is information about the instrument.

Ariadne's Thread builds abstention into the definition of its output. Each posture is replicated to estimate within-posture variance. The between-posture signal is tested against that noise with a standard variance-ratio test, and Z is reported only when the signal clears the threshold at a declared level. Below it, the honest output is an abstention with its reason, not a small number that looks like a finding.

Abstention is only as good as the rules around it. The test is conservative at small panels, so sensitivity has to be bought with replicates. Reporting only when the test passes inflates values near the threshold, so the abstention rule and its level are published with every study. Z is given as an interval, never a bare point, and a Z reported without its versioned panel declaration, noise floor and abstention status is treated as a rhetorical device rather than a measurement.

The principle reaches beyond Z. A legal output that cannot say "the evidence here does not support a reading" will fill the silence with confidence it does not have. Any tool that measures something about the law, whether openness, strength or risk, should publish when it declines to answer, how often, and why. A refusal rate is not an embarrassment to be hidden. It is part of the instrument's specification.