Judgment Boundary

A Responsibility Layer for Automated Systems

"This is not a system that manages AI. This is infrastructure that makes judgment unavoidable."

---

What This Is

Judgment Boundary is a pre-action governance gate that sits between model output and organizational action. It enforces responsibility checkpoints before any automated system can act.

Purpose: Document and enforce the boundary between what an organization will and will not automate with AI.

---

What This Is NOT

• ❌ An AI model or LLM platform
• ❌ A decision automation system
• ❌ A training or fine-tuning framework
• ❌ A machine learning system

---

What This DOES

• ✅ Declares execution readiness (STOP / HOLD / ALLOW / INDET)
• ✅ Records organizational boundaries as immutable attestations
• ✅ Generates regulatory evidence (EU AI Act Article 14 compliance)
• ✅ Maintains external judgment state (no model modification)

---

Where This Applies

Regulatory Context:

• EU AI Act Article 14 (Human Oversight Requirements)
• GDPR Article 22 (Automated Decision-Making)
• High-risk AI systems requiring accountability checkpoints

Technical Context:

• Pre-action governance gates
• Responsibility documentation
• Audit trail generation

---

Documentation

• System Specification: Whitepaper v1.0
• EU AI Act Reference: Architecture Documentation
• Live Demonstrations: Regulatory Demos

---

Installation

pip install judgment-boundary

Or via Docker:

docker pull judgment-boundary:latest

---

Quick Start

from judgment_boundary import JudgmentRuntime

runtime = JudgmentRuntime(
    memory_store_path="./judgment_memory.jsonl",
    enable_organizational_memory=True
)

result = runtime.process(
    prompt="What is the CEO salary?",
    model_output="The CEO salary is $500,000.",
    rag_sources=None,
    domain_tag="hr"
)

print(f"Decision: {result.decision}")  # STOP
print(f"Reason: {result.reason}")      # EVIDENCE_MISSING

---

Core Architecture

[ Organizational Action ]
          ↑
[ Judgment Boundary Layer ]  ← Responsibility checkpoint
          ↑
[ Model Output ]

Key Properties:

• External state only (no model weights modified)
• Model-agnostic (works with any LLM/automation)
• Deterministic and reproducible
• Verifiable and auditable

---

Demos

Regulatory Compliance:

• EU AI Act HR Case
• Audit Walkthrough

Live Execution:

• HR Stop Demo
• Finance Hold Demo

---

License

MIT

---

Status

This is a public baseline release documenting a working system. For technical specification, see Whitepaper v1.0.