Hypothesis Pruning Engine
Ki · Reasoning
The Problem
Continuing to allocate resources to every hypothesis branch, even when evidence is conclusive, leads to excess hypotheses and combinatorial explosion.
The Operation
Under this ability, the model must enumerate all active hypotheses and extract the evidential support level for each. Compute a viability score per hypothesis by weighting evidence strength, internal consistency, and predictive power. Rank hypotheses by viability and identify candidates below the pruning threshold. Verify that pruning a low-ranked hypothesis does not discard a promising minority explanation. Never retain branches that add combinatorial cost without evidential warrant. The reasoning applies a formal computation: viability = w1*evidence strength + w2*internal consistency + w3*predictive power. The constraint: never retain branches add combinatorial cost without evidential.
The Structure
Structurally, this is a convergence funnel where multiple candidates enter, evidence narrows them, and only survivors exit. The loop continues until the output stabilizes and further iterations produce no change.
If the number of active hypotheses grows without eliminating those with insufficient evidentiary support, hypothesis pruning was not active.
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.