Masked L1 distance is the primary similarity metric in AGISystem2. It measures the Manhattan distance between two vectors, but only considers dimensions marked as relevant by a mask. This enables context-sensitive comparisons where different dimensions matter in different situations.

Why Masking?

Not all dimensions are relevant in every context. When comparing animals by behavior, physical appearance dimensions might be irrelevant. When comparing products by price, brand dimensions might not matter. The mask provides this flexibility without requiring different vector representations for different contexts.

AGISystem2 uses separate dimension ranges for ontological (factual) and axiological (value) information. The BiasController can automatically mask value dimensions when pure factual comparisons are needed, enabling the "veil of ignorance" mode for fair reasoning.

Implementation

The MathEngine provides optimized distance computation:

// Basic masked L1 distance
const dist = mathEngine.maskedL1(vectorA, vectorB, mask);

// With ontology-only mask (dims 0-255)
const ontoDist = mathEngine.maskedL1(vectorA, vectorB, ontologyMask);

// With full mask (all 384 dims)
const fullDist = mathEngine.maskedL1(vectorA, vectorB, fullMask);

Mask Types

Mask Type	Dimensions	Use Case
Full mask	All 384	Complete similarity including values
Ontology mask	0-255	Factual similarity only (veil of ignorance)
Axiology mask	256-383	Value similarity only
Custom mask	User-defined	Domain-specific comparisons

Why L1 (Manhattan)?

L1 distance sums absolute differences rather than squared differences (L2/Euclidean). This has practical advantages:

• Int8 friendly – No multiplication needed, avoiding overflow with saturating addition
• Sparse-aware – A single large difference dominates, matching intuition about key features
• Closed under diamonds – L1 balls (diamonds) have clean geometric properties

Related Documentation

• MathEngine – Implementation details
• Bounded Diamonds – How masks integrate with concept shapes
• BiasController – Automatic masking for fair reasoning
• Ontology – Factual dimension space
• Axiology – Value dimension space