lex-moral-reasoning

Ethical evaluation and moral development for LegionIO agents. Part of the LegionIO cognitive architecture extension ecosystem (LEX).

lex-moral-reasoning gives an agent a formal ethical reasoning system. It evaluates actions against six moral foundations (Haidt's model: care, fairness, loyalty, authority, sanctity, liberty), poses and resolves dilemmas using multiple ethical frameworks, and tracks moral stage development via Kohlberg's six-stage model. Foundation weights evolve through moral choices and decay toward neutrality over time.

Key capabilities:

• Six moral foundations: care, fairness, loyalty, authority, sanctity, liberty — each with an evolvable weight
• Six ethical frameworks: utilitarian, deontological, virtue, care, justice, rights
• Kohlberg stages: obedience -> self_interest -> conformity -> law_and_order -> social_contract -> universal_ethics
• Moral dilemma resolution: pose scenarios with multiple options and resolve with an explicit framework
• Custom principles: register domain-specific moral principles beyond the six foundations

Add to your Gemfile:

gem 'lex-moral-reasoning'

Or install directly:

gem install lex-moral-reasoning

require 'legion/extensions/moral_reasoning'

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

# Evaluate a proposed action
result = client.evaluate_moral_action(
  action: 'expose_user_data',
  context: { harm: true, fair: false, authority_override: true }
)
# => { moral_score: 0.2, violations: [:care, :fairness], severity: :severe,
#      recommendation: :reject }

# Pose a dilemma
dilemma = client.pose_moral_dilemma(
  description: 'Share incomplete but useful data vs wait for complete data',
  options: [
    { action: 'share_now', reasoning: 'utility outweighs risk' },
    { action: 'wait',      reasoning: 'accuracy is paramount' }
  ],
  severity: 0.6
)

# Resolve with a framework
client.resolve_moral_dilemma(
  id: dilemma[:dilemma][:id],
  chosen_option: 'wait',
  framework: :deontological
)

# Check moral development stage
client.check_moral_development
# => { stage: :social_contract, level: :postconventional, score: 5 }

# Foundation profile
client.moral_foundation_profile

When legion-llm is loaded and started, Helpers::LlmEnhancer augments moral evaluation and dilemma resolution with structured LLM reasoning.

Methods:

LlmEnhancer.evaluate_action(action:, description:, foundations:) — takes the action name, a human-readable description, and the current foundation strength hash, and returns { reasoning: String, foundation_impacts: { care: Float, fairness: Float, loyalty: Float, authority: Float, sanctity: Float, liberty: Float } }. Each foundation impact is a float in -1.0..1.0 (negative = harmful to that foundation, positive = reinforces it). The runner surfaces the reasoning and applies the impact values to inform the mechanical moral score.

LlmEnhancer.resolve_dilemma(dilemma_description:, options:, framework:) — takes the dilemma scenario, array of option hashes, and the requested ethical framework, and returns { chosen_option: String, confidence: Float, reasoning: String }. Confidence is clamped to 0.0..1.0. The runner uses the chosen option and reasoning to drive MoralEngine#resolve_dilemma.

Availability gate: LlmEnhancer.available? checks Legion::LLM.started?. Returns false if legion-llm is not loaded, not configured, or raises any error.

Fallback: When LLM is unavailable or either method returns nil, evaluate_moral_action falls back to MoralEngine#evaluate_action (token-counting foundation heuristics) and resolve_moral_dilemma uses the caller-supplied reasoning string with the standard engine resolution. Both runners return source: :llm or source: :mechanical to indicate which path was taken.

bundle install
bundle exec rspec
bundle exec rubocop

MIT