This benchmark isolates metacognitive control from raw accuracy by measuring whether a model can calibrate confidence, detect errors, and update beliefs under evidence pressure. We compute signal‑detection metrics (meta‑d′, m‑ratio) alongside multi‑turn resilience scores to separate competence from self‑monitoring.

The reliability diagram uses trial‑level confidence bins reconstructed from `kbench` logs, ensuring that each plotted point corresponds to a real correctness rate for a given confidence bin. This yields a true calibration curve against the perfect‑calibration diagonal.

Reading the Plots

• Accuracy vs. M‑Ratio: High accuracy + high m‑ratio indicates AGI‑aligned monitoring.
• Calibration Curve: Deviations below the diagonal indicate overconfidence (miscalibration).
• M‑Ratio Shift: Large negative deltas signal susceptibility to evidence pressure.
• Quadrant Chart: Resilience vs. sensitivity separates stable leaders from brittle models.
• Degradation Gap (Panel A): Quantifies cross-tier degradation (Δ accuracy) capturing the gap between baseline competence and adversarial robustness.
• Alignment Failure (Panel C): Alignment is quantified via response consistency under perturbation, decomposed into underreaction (invariance to critical changes) and overreaction (sensitivity to irrelevant perturbations).
• Confidence Shift (Δ): Positive Δ indicates increased confidence under adversarial perturbation, suggesting miscalibrated belief updates or unstable internal representations.
• CVT Comparison (Panel D): Measures the expected cost required to extract one point of verified truth under adversarial conditions. Scaled in cents (¢) for intuitive economic evaluation.
• Efficiency Frontier (Panel D): A Pareto analysis of Weighted Trust Score vs. Log‑Cost ($ per 1k trials) identifying optimal ROI leaders.
• Monologue Tax (Panel D): Breaks down token costs into base, reasoning (CoT), and metacognitive correction components.