lex-bias
Cognitive bias detection and correction for brain-modeled agentic AI.
What It Does
Detects five cognitive biases in the agent's decision-making contexts and computes correction factors. Tracks per-bias susceptibility over time via EMA. The five biases modeled are: anchoring (initial values distorting estimates), confirmation (over-weighting supporting evidence), availability (distortion from memorable recent events), recency (over-weighting the most recent data), and sunk cost (continuing failing investments to justify past costs).
Core Concept: Bias Detection Against Decision Context
# Provide context for bias detection
result = client.check_for_bias(
domain: :budget,
decision_context: {
current_value: 150_000,
evidence_direction: :positive,
hypothesis_direction: :positive, # same direction → confirmation bias check
recent_events: [:cost_overrun, :cost_overrun, :minor_savings], # availability
data_points: [0.9, 0.8, 0.7, 0.3, 0.2], # recency (0.2/0.3 most recent)
invested: 500_000,
expected_return: 0.1 # low ROI → sunk cost check
}
)
# => { detected: [{ bias_type: :sunk_cost, magnitude: 0.7, correction_applied: 0.35 }] }Usage
client = Legion::Extensions::Bias::Client.new
# Record anchors for anchoring bias detection
client.record_anchor(domain: :performance, value: 100.0)
# Check for active biases before a decision
client.check_for_bias(domain: :performance, decision_context: { current_value: 95.0 })
# View susceptibility profile
client.susceptibility_profile
# => { anchoring: 0.6, confirmation: 0.4, availability: 0.5, recency: 0.3, sunk_cost: 0.4 }
# Maintenance (anchor decay runs automatically via Update actor)
client.update_biasIntegration
Call check_for_bias before high-stakes decisions in lex-tick's action_selection phase. Pairs with lex-anchoring (shares anchor data), lex-appraisal (provides emotional/relevance context for confirmation bias), and lex-anosognosia (bias susceptibility as judgment deficits).
Development
bundle install
bundle exec rspec
bundle exec rubocopLicense
MIT