Auditing the exported p value of a neural information detector

A detector's descriptive score and its reported significance need separate validation. I audit the standalone neural mutual-information p-value exported by a pinned implementation of Audit the Whisper, a detector reused in a subsequent collusion study. Its nominal-null fits retain the observed joint sample while changing only the product-reference sample, and use 120 optimization steps; the observed fit uses 200. An exact constructed example shows that this reference can retain the dependence being tested. A separate, predeclared neural mechanism experiment varies joint construction and training budget across 40 synthetic datasets. Under correlation 0.9, equal-budget reference means are 0.52703 nats when the joint is retained and 0.01686 when it is permuted. These values describe a local implementation. Operational performance and Gaussian-tail calibration from three fits per run pair require separate evaluation. I propose a local output-only quarantine that preserves diagnostic scores while marking inferential output unavailable. A bounded synthetic execution of the pinned detector and estimator with local adapters preserves scores under the patch; full-pipeline integration and calibration remain untested. A source trace distinguishes a stress-script consumer of the exported value from the default union, which excludes this detector, and the raw-score ROC and externally derived honest-calibration operating thresholds. The audit provides a source-level diagnosis and regression checks for the proposed quarantine. A calibrated replacement test remains open.