jac-agent

JAC: Judgment Accountability Chain — Official Reference Implementation

A minimal, high-performance, verifiable decision audit and training layer for AI agents, fully compliant with IETF Internet-Draft specifications.

---

Compliance

This implementation strictly adheres to the following specifications:

• draft-wang-jep-judgment-event-protocol-04 (Judgment Event Protocol)
• draft-wang-hjs-accountability-04 (Accountability Layer for AI Agents)
• draft-wang-jac-01 (Judgment Accountability Chain)

No custom fields, no speculative logic, no unverified extensions.

---

Core Features

1. Standard Audit Mode

• Immutable decision event chains
• Cryptographic verifiability (RFC 8785, RFC 9122)
• Privacy-preserving identity (digest-only anonymity)
• Full audit report export
• Cross-agent traceability via ref and task_based_on

2. High-Performance Training Mode

• In-memory batching (no I/O blocking during training)
• Thread-safe, low-latency event recording
• Automatic causal chain slicing for SFT / RL / DPO training
• High-throughput support for millions of agent decisions
• Zero overhead to agent inference or training loops

---

Installation

pip install jac-agent

---

Quick Start

Standard Usage (Audit & Compliance)

from jac_agent_trace import judge, show_trace_chain, export_audit_report

judge(
    subject="User request validation",
    judgment="Approve legitimate user action",
    evidence="Input verified against policy",
    risk_level="low"
)

show_trace_chain()
export_audit_report()

High-Performance Agent Training

from jac_agent_trace import (
    judge,
    enable_training_mode,
    export_training_dataset,
    export_audit_report
)

# Enable high-throughput training mode
enable_training_mode(batch_size=32)

# Record thousands of decisions without blocking
for i in range(1000):
    judge(
        subject=f"Task-{i}",
        judgment=f"Decision-{i}",
        evidence=f"Observation-{i}",
        risk_level="low"
    )

# Export causal chain dataset for direct training
export_training_dataset()

# Export full audit trail
export_audit_report()

---

Performance

• Hash computation: <0.1ms per event
• Batch memory buffering: no disk I/O during training
• Causal chain slicing: O(n) structured training data generation
• Supports millions of events in continuous agent operation

---

Protocol Stack

1. JEP: Core event format, verbs (J/D/V/T), nonce, signature
2. HJS: Accountability, privacy, risk levels, immutable machine records
3. JAC: Causal decision linking via task_based_on

---

Use Cases

• AI Agent decision traceability
• Regulatory compliance & auditing
• Reinforcement Learning (RL) causal trajectory logging
• Supervised Fine-Tuning (SFT) dataset construction
• Multi-agent collaboration verification
• Fault and hallucination root-cause analysis

---

License

Apache 2.0 license