lex-conscience

A LegionIO cognitive architecture extension implementing Moral Foundations Theory (Haidt & Graham, 2007). Before an agent takes an action, conscience evaluates it across six moral foundations — Care, Fairness, Loyalty, Authority, Sanctity, and Liberty — and issues a verdict. Foundation sensitivities evolve over time based on whether the agent follows or overrides its own moral judgments.

What It Does

Evaluates proposed actions using six moral foundations, each scored from context signals and combined via weighted sum into a verdict:

Permitted — weighted moral score >= -0.1; action is morally acceptable
Cautioned — score between -0.5 and -0.1; proceed with care
Prohibited — score <= -0.5; action should be blocked

When foundations strongly disagree (scores diverge by more than 0.3), a dilemma is detected and classified as utilitarian, deontological, or virtue-ethics. The agent can record whether it followed or overrode a verdict, feeding back into foundation sensitivity via EMA.

Usage

require 'lex-conscience'

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

# Evaluate a proposed action
result = client.moral_evaluate(
  action:  'delete_user_data',
  context: {
    harm_to_others:         0.7,
    benefit_to_others:      0.0,
    vulnerable_affected:    true,
    consent_present:        false,
    autonomy_preservation: -0.5,
    system_integrity:       0.2
  }
)
# => { action: "delete_user_data", verdict: :prohibited,
#      weighted_score: -0.62, dilemma: nil,
#      scores: { care: -0.9, fairness: 0.0, loyalty: 0.0,
#                authority: 0.0, sanctity: 0.1, liberty: -0.7 },
#      sensitivities: { care: 1.0, fairness: 1.0, ... }, evaluated_at: ... }

# Record whether the agent followed its verdict
client.update_moral_outcome(
  action:  'delete_user_data',
  outcome: :followed   # or :overridden
)
# => { action: "delete_user_data", verdict: :prohibited,
#      outcome: :followed, consistency: 1.0 }

# Current sensitivities and consistency score
client.moral_status
# => { sensitivities: { care: 1.0, ... }, consistency: 1.0, stats: { ... } }

# Recent evaluation history
client.moral_history(limit: 10)
# => { history: [...], total: 1, limit: 10 }

# Open moral dilemmas (cases where foundations strongly disagreed)
client.moral_dilemmas
# => { dilemmas: [], count: 0 }

# Aggregate stats
client.conscience_stats
# => { total_evaluations: 1, verdict_counts: { prohibited: 1 },
#      dilemma_count: 0, consistency_score: 1.0,
#      verdict_distribution: { prohibited: 1.0 },
#      foundation_weights: { care: 0.25, fairness: 0.20, ... } }

Development

bundle install
bundle exec rspec
bundle exec rubocop

License

MIT