Bias control in AGISystem2 uses dimension masking to prevent protected attributes from influencing decisions. The BiasController can selectively disable dimensions during reasoning, implementing a geometric "veil of ignorance".

How It Works

Bias control operates by modifying the masked L1 distance calculation:

1. Define protected dimensions – Specify which dimensions encode sensitive attributes
2. Create mask – Set protected dimension weights to 0
3. Apply during reasoning – All similarity and containment checks use the masked distance
4. Log decisions – AuditLog records when bias masks were applied

Mask Types

Mask Type	Dimensions Masked	Use Case
Ontology-only	256-383 (all axiology)	Pure factual reasoning
Protected attributes	Specific ontological dims	Fair hiring, lending
Custom domain	Domain-specific subset	Medical, legal contexts
Full mask	None (all dims active)	Normal reasoning

Implementation

// Define protected dimensions
biasController.setProtectedDimensions([120, 121, 122, 123]); // e.g., age, gender

// Create mask (384 dims, 1 = active, 0 = masked)
const mask = biasController.createFairnessMask();
// Result: [1,1,...,1,0,0,0,0,1,1,...] with 0s at positions 120-123

// Use in reasoning
const fairDistance = mathEngine.maskedL1(conceptA, conceptB, mask);

// Audit log entry created automatically:
// { action: "BIAS_MASK_APPLIED", dims: [120-123], context: "employment_decision" }

Veil of Ignorance Mode

For maximum fairness, enable veil of ignorance mode which masks all axiological dimensions plus protected ontological ones:

// Full veil: no values, no protected attributes
biasController.enableVeilOfIgnorance({
  protectedOntology: [120, 121, 122, 123],
  includeAllAxiology: true
});

// Now all reasoning ignores both values and protected attributes
ASK "Should applicant be hired?" # Based only on qualifications

Bias Detection

The system can detect potential bias by comparing results with and without masks:

// Run query twice: with and without mask
const normalResult = reasoner.ask(query, { mask: fullMask });
const fairResult = reasoner.ask(query, { mask: fairnessMask });

// If results differ, bias may be present
if (normalResult !== fairResult) {
  auditLog.warn("BIAS_DETECTED", { query, diff: [...] });
}

Related Documentation

• BiasController – Implementation details
• Veil of Ignorance – Full fairness mode
• Axiology – Value dimensions (often masked)
• Ontology – Factual dimensions (selectively masked)
• Auditability – Logging bias control decisions
• Masked L1 Distance – The underlying metric