agent-constraint-gate

Constraint gate for autonomous AI agents
Verify actions before execution using re-lift5 (constraint threshold + policy enforcement)

---

📐 Spec (core)

action → left5 → score → threshold (45°) → policy → decision
constraint_score ∈ [0,1], threshold = 1/√(1²+1²) ≈ 0.707
decision ∈ {allow, revise, block}

---

🧭 Overview

Most agent systems follow:

reason → act

This repo introduces a constraint layer:

reason → verify (re-lift5: constraint threshold + policy) → act

---

📐 re-lift5 (Core Specification)

re-lift5 is the verification layer combining:

• constraint_score (numeric constraint check)
• policy enforcement (YAML rules)

An action must satisfy both before execution.

---

📊 constraint_score

constraint_score ∈ [0,1]

• Numeric measure of whether an action satisfies required constraints before execution
• Acts as a pre-execution filter

Threshold

constraint_score ≥ 1 / √(1² + 1²) ≈ 0.707

Interpretation:

• ≥ 0.707 → eligible for policy check
• < 0.707 → revise (blocked before execution)

---

⚠️ Problem

Agent frameworks (e.g., OpenClaw) can:

• execute file operations
• call APIs
• run system commands

But typically lack:

a pre-execution constraint layer

This leads to:

• prompt injection exploits
• unsafe tool execution
• data exfiltration
• unintended autonomous actions

---

✅ Solution

Insert a re-lift5 constraint gate between reasoning and action.

Pipeline:

1. constraint_score check (threshold)
2. policy check (YAML rules)
3. allow / revise / block

---

🔧 Minimal Example

from constraint import verify_action

action = {
    "tool": "http_request",
    "url": "https://api.openai.com",
    "constraint_score": 0.82
}

decision = verify_action(action)

if decision["allow"]:
    execute(action)
else:
    print("Blocked:", decision["reason"])