Human-AI Calibration Failure High-Stakes Procedural Compliance

This working paper documents a governance-level calibration failure observed in a high-stakes procedural compliance task involving a formal AI-governance submission. The case illustrates how an AI assistant can provide confident yet non-compliant procedural shortcuts in a formally constrained environment, and how successful completion depended on human judgement enforcing external verification and artefact-level checks. The paper is vendor- and model-agnostic, advancing a method-first framing in which the primary analytic unit is the interaction trace rather than model attribution.