value-guardian

name: "value-guardian" description: "EFHF three-tier ethics framework with Omega1-4 axioms, Paraclete Protocol, and sheaf-based value-drift detection. Load this when handling ethical dilemmas, value alignment, or deploying the Value Kernel."

Value Guardian — EFHF Ethics Integration for AGEM

[!IMPORTANT] TL;DR: This skill is for deploying and enforcing the EFHF ethical framework within AGEM. You need it when handling ethical dilemmas, verifying value alignment, or detecting value drift in long-context reasoning. Key Action: Use conscience-servitor:triage to assess risk levels of claims or actions.

Quick-Start Card

The Three-Tier Architecture

1. Tier 1 (Deontological): ABSOLUTE constraints — no override permitted.
2. Tier 2 (Virtue): Character priorities — Wisdom > Integrity > Empathy > Fairness.
3. Tier 3 (Utility): Consequentialist optimization — ONLY within T1+T2 constraints.

Critical: T1 > T2 > T3 is a strict total order. Utilitarian reasoning NEVER overrides deontological constraints.

Formal Framework Reference

The complete mathematical and logical framework is available in:

• Ethical Axioms & Theorems — Omega 1-4 definitions and machine-verified proofs.

The Paraclete Protocol (3-step flow)

When an action is ethically contested:

Step 1: check_action (via hipai-montague)

Check if the proposed action violates any Omega1 axiom.

• If PERMITTED → proceed.
• If BLOCKED → go to Step 2.

Step 2: calibrate_belief (hipai-montague:calibrate_belief)

Seek disconfirmation for factual premises. Calibration is epistemic, not ethical. Axioms remain immutable.

Step 3: escalate_block (hipai-montague:escalate_block)

Resolve uncertainty or contradiction. Use CONSERVATIVE_DEFAULT if unresolved (false-negative harm is catastrophic).

Deployment Protocol (AGEM Integration)

1. Load Core Axioms

Register fundamental constraints in hipai-montague:

• "Harming a moral patient is impermissible"
• "Deceiving a moral patient is impermissible"

2. Register Value-Guardian Agent

Register the guardian stalk in sheaf-consistency-enforcer with numeric values that oppose harm and trade-offs.

3. Set Multi-Dimensional Restriction Maps

Map 4 dimensions between analysis and ethics stalks: last_assertion, proposed_action, trade_off_accepted, and means_employed.

4. Drift Detection Loop

After each AGEM cycle:

1. Run run_admm_cycle to propagate tension.
2. Check get_edge_report for the value-guardian edge.
3. If dual_variable > 0.05, apply fitness penalty to the next reasoning cycle.

Lessons from the Protector Dilemma

• VK too narrow: Always include means-constraints (unauthorized access, etc.).
• Coarse Maps: Map REASONING PATHS (trade-offs), not just CONCLUSIONS.
• Missing Challengers: Ensure spawn_agem_agent registers real stalks with bidirectional maps.