MCMetacognitionID: MC-009

Performance Meta-Evaluator

Ki · Reasoning

The Problem

Eight reasoning steps, each exploring a different angle. Uniform progress feels like the analysis is on track. But progress without convergence is just motion.

The Operation

This ability makes the model assign a composite score, advancing, marginal, or redundant. Flag performance blindness: if two consecutive steps score redundant, reasoning is going in circles or spinning wheels. Apply EMA weight updates, recent step quality determines strategy weighting over early performance. If a marginal step is detected, test it against a counterexample before accepting. Repeat Steps 1-4 until convergence. The reasoning applies a formal computation: step score = EMA(alpha, [advancing=1, marginal=0.5, redundant=0]). The constraint: never apply uniform weighting regardless quality.

The Structure

Under the hood, the reasoning follows an iterative convergence loop that cycles until the reasoning stabilizes on a consistent answer. The monitor runs continuously, checking for drift at each step.

If successive reasoning steps restate the same content without advancing the conclusion, step quality evaluation was not applied.

Haki · Reasoning-Multi

Cross-Domain Suppression

In Haki mode, the API retrieves the primary ability first, then fans out to three synergy roles that compound its reasoning.

When retrieved in Haki mode, the primary ability is augmented with failure guards extracted from 3 abilities in different cognitive domains. Each guard blocks a specific reasoning failure the primary alone wouldn't catch. A self-check forces verification before output. The result is cross-domain coverage that no single ability can reach alone.