The Withheld-Metric Condition: Uncertainty Field Adequacy and the Limits of Mandatory AI Transparency

This is a translational article examining a specific blind spot in mandatory AI transparency schemas, drawn from a structured audit of the UK's Algorithmic Transparency Recording Standard (ATRS). Mandatory AI transparency frameworks convert public-sector algorithmic tool use into structured public data, but formal compliance with schema requirements does not guarantee that records preserve the information needed for meaningful accountability. This paper identifies the withheld-metric condition: a performance metric that would let an affected person assess an AI tool's reliability exists and is operationally active, yet is absent from the public record — and that absence is, under current schema design, indistinguishable from the absence produced by never having measured at all. A structured audit of 46 high-stakes ATRS records shows the uncertainty field absent in 29 records (63%), that no record achieves full event-level uncertainty disclosure, and that the audit instrument itself was forced to encode the conflation: strict scoring treats deliberate withholding identically to unmeasured absence. The remaining seventeen records reach at most generic acknowledgement, including the corpus's own best-practice record. Working from three cases that occupy three distinct epistemic states, the paper proposes a four-element field-content adequacy standard — category acknowledged, metric named, value or threshold published, justified redaction where constrained — and sets out what DSIT, EU AI Act implementers, and the G7 Hiroshima Process would each need to change to close this gap.