Model Critic
Ki · Reasoning
The Problem
The model's accuracy is high, so we trust its predictions without questioning potential algorithmic bias or training data limitations.
The Operation
When activated, the model must extract the model's core assumptions, identify training data scope, feature dependencies, and calibration method. Scan for data leakage, distributional shift, and confounded features. Do not trust vendor accuracy claims without independent validation on held-out data. Verify that confidence scores are calibrated by testing predicted probabilities against observed frequencies. Simulate adversarial inputs to probe boundary conditions. If it detects trust vendor accuracy claims without independent validation, it halts and corrects.
The Structure
This ability runs on an adversarial self-attack that actively tries to break its own conclusions.
If a model's output is accepted without questioning its training assumptions or data biases, adversarial model evaluation has failed.
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.