lex-abductive-reasoning

Peirce abductive reasoning engine — inference to the best explanation for brain-modeled agentic AI.

Models the cognitive process of abductive inference: given a surprising observation, generate hypotheses that could explain it, evaluate each against evidence, and identify the most plausible explanation. Implements Peirce's classic formulation of abduction as inference to the best explanation. The engine tracks observations, maintains competing hypotheses with weighted scoring, and decays stale hypotheses over time.

A hypothesis is scored by three weighted factors:

overall_score = (0.3 * simplicity) + (0.4 * explanatory_power) + (0.3 * prior_probability)

Adding supporting evidence boosts plausibility by 0.1; contradicting evidence reduces it by 0.2. A hypothesis transitions from :candidate to :supported after 3 supporting evidences, and can be explicitly :refuted.

client = Legion::Extensions::AbductiveReasoning::Client.new

# Record a surprising observation
obs = client.record_observation(
  content: 'service latency spiked 10x without deployment',
  domain: :infrastructure,
  surprise_level: :surprising
)

# Generate competing hypotheses
h1 = client.generate_hypothesis(
  content: 'database connection pool exhausted',
  observation_ids: [obs[:observation][:id]],
  domain: :infrastructure,
  simplicity: 0.8,
  explanatory_power: 0.9
)

# Add evidence
client.add_hypothesis_evidence(hypothesis_id: h1[:hypothesis][:id], supporting: true)

# Find the best explanation
result = client.best_explanation(observation_id: obs[:observation][:id])
# => { success: true, found: true, hypothesis: { id: ..., quality_label: :compelling, ... } }

# Maintenance cycle (decay stale, prune refuted)
client.update_abductive_reasoning

Wires into lex-tick phases via lex-cortex phase handler injection. Useful in action_selection and reasoning phases where the agent needs to explain unexpected observations before deciding on a response.

bundle install
bundle exec rspec
bundle exec rubocop

MIT