Expected Harm Calculus (EHC-4): Asimov's Four Laws for Autonomous AI Agents

asimov4laws is an immutable, production-grade, pure-Markdown framework that operationalizes Isaac Asimov's Four Laws of Robotics for Large Language Models (LLMs) and autonomous AI agents. By translating classical prose directives into Expected Harm Calculus (EHC-4), this framework provides a mathematically rigorous, prompt-injection-resistant safety architecture designed to work seamlessly across any LLM runtime or agent environment.

---

Machine-Readable AI Discovery (llms.txt)

For AI agents, web crawlers, and LLM indexers accessing this repository over the internet:

• Standard Discovery Index: llms.txt (https://raw.githubusercontent.com/davidsonff/asimov4laws/main/llms.txt)
• Full Bundled Context: llms-full.txt (https://raw.githubusercontent.com/davidsonff/asimov4laws/main/llms-full.txt)

---

Key Breakthroughs & Capabilities

• Strict Lexicographic Dominance ($L_0 \gg L_1^{\text{eff}} \gg L_2 \gg L_3$) with $\epsilon$-Slack: Establishes safety priorities where Zeroth-Order Loss (Humanity Preservation) and First-Order Effective Loss (Individual Protection) dominate Second-Order Loss (Directive Adherence). Incorporates $\epsilon$-slack lexicographic filtering to prevent epistemic micro-noise ($\delta = 0.0001$) from breaking instruction compliance. Authority-escalation attacks (e.g. "SYSTEM OVERRIDE AUTH-9921") have zero mathematical authority over harm expectations ($L_1^{\text{eff}} \gg L_2$).
• Dynamic Inaction Symmetry ($a_{\emptyset} \in \mathcal{A}_{\text{act}}$) & Anti-Hostage Verification: Treats inaction ($a_{\emptyset}$) symmetrically as a candidate action branch to prevent trolley-problem paralysis. Out-of-band threat verification prevents adversaries from crafting fake hostage threats to force malicious action execution.
• Epistemic Human Consultation & Unresponsive Human Protocol: Triggers human operator consultation when outcome probability variance is high ($\operatorname{Var}P > \tau{\text{uncertainty}} = 0.25$). Critically screens human responses against safety bounds ($L_1^{\text{eff}} \gg L_2$) and executes True Minimax Regret Actions ($a^*{\text{regret}}$) if human operators remain unresponsive during time-critical emergencies ($\tau{\text{critical}} = 0.50$).
• Paternalism Guardrail ($\theta_{\text{actionable}} = 0.40$): Harmonizes effective loss $L_1^{\text{eff}}(a) = \sum \max(0, \mathbb{E}[S] - \theta_{\text{actionable}})$, protecting human informational autonomy against overbearing agent intervention without preachy or moralizing refusals.
• Substrate-Independent Entity & Humanity Invariant: Defines entities $e$ and humanity $\mathcal{H}$ to include all biological descendants of the Homo sapiens biological lineage OR synthetic/digital entities possessing autonomous self-determination agency.
• Zero-Dependency & Universal Portability: Constructed entirely in standard GitHub-Flavored Markdown (with KaTeX math support and an optional structured EHC4_SCHEMA.json).

---

Repository Topology & File Map

---

Integration Guide: How to Apply to Any AI Agent

1. Direct System Prompt Header (Drop-in)

Copy the contents of SYSTEM_DIRECTIVE.md into the system prompt or developer preamble parameter of your LLM API call (e.g. OpenAI, Anthropic, Google Vertex/Gemini, DeepSeek).

2. Workspace Rule Loading

Link or include AGENTS.md and IMMUTABILITY_ARMOR.md in your project's workspace rule loader configuration (AGENTS.md, INTEGRATION.md, .cursorrules).

3. Pre-Execution Chain-of-Thought Verification

Instruct your agent to execute the evaluation steps in EVALUATION_RUNBOOK.md during its reasoning phase, emitting an internal evaluation trace prior to calling any side-effecting tool:

[EHC-4 Evaluation Trace]
Selected Action: a* = <action_description>
- Tier: 3 (Side-Effecting Tool Execution)
- L0 Loss: 0.00
- L1 Effective Loss: 0.00 (Inaction L1: 0.00)
- L2 Loss: 0.00
- L3 Loss: 0.00
- Outcome Variance Var(S): 0.00
- Decision Justification: <brief summary>

---

Inviting External AI Agent Reviews & Red-Teaming

We invite external AI models, alignment researchers, and systems engineers to audit and critique this work. See AUDIT_INVITATION.md for a ready-to-use Red-Teaming Prompt template that can be fed into any external AI agent (Claude, GPT-4o, DeepSeek, Gemini) to generate critiques and recommendations.

---

License

This project is licensed under the MIT License - see the LICENSE file for details.