Salience Gradient Orchestrator
Ki · Reasoning
The Problem
The retrieval system ranks documents based solely on proximity, leading to flat priority and static relevance scores.
The Operation
When activated, the model must compute the salience gradient across the full conceptual space relative to the current query attractor. Rank concept clusters by gradient magnitude and trace steepest-ascent paths. Identify gradient plateaus where salience is uniform and flag them as ambiguity zones. Verify that retrieval follows gradient direction, not mere proximity. If the attractor shifts mid-analysis, recompute the gradient from the new anchor. The reasoning applies a formal computation: salience gradient = d(salience) / d(conceptual distance from attractor). If it detects route retrieval by proximity alone, it halts and corrects.
The Structure
The reasoning structure is a convergence funnel where multiple candidates enter, evidence narrows them, and only survivors exit. Execution repeats until the reasoning locks onto a stable conclusion.
If all retrieved information is weighted equally without a salience gradient prioritizing the most relevant elements, salience orchestration 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.