Skip to content

aiuc-1-evidence.md

Latest commit

 

History

History
72 lines (50 loc) · 5.15 KB

aiuc-1-evidence.md

File metadata and controls

72 lines (50 loc) · 5.15 KB
Raw

EPI AIUC-1 Compliance & Evidence Specification

Status: Active
Date: 2026-05-18
Version: 1.0.0
Authors: EPI Project Team

Abstract

This document specifies how the Evidence Packaged Infrastructure (EPI) framework acts as the definitive technical evidence container for the AIUC-1 Compliance and Assurance Framework (commonly referred to as the "SOC 2 for AI agents"). Specifically, it maps EPI’s cryptographic, timeline, and policy structures directly to the six core trust domains evaluated under an AIUC-1 audit.

1. Overview of AIUC-1

The AIUC-1 framework is the industry-standard compliance and assurance program designed specifically for autonomous AI agents. Unlike broad governance structures that only evaluate corporate policy, AIUC-1 demands verifiable technical proof of an AI agent's behavior, safety limits, and boundaries in production, including mandatory quarterly adversarial testing and continuous operational audit logging.

By packing all agent inputs, outputs, environmental parameters, and evaluation records into a single cryptographically sealed .epi file, EPI provides a portable, self-contained evidence container that auditors can verify offline to demonstrate compliance with AIUC-1 controls.

2. Core AIUC-1 Trust Domains & EPI Mapping

EPI satisfies the strict evidence requirements of the six AIUC-1 trust domains:

2.1 Security (Control Domain 1)

  • AIUC-1 Requirement: Verifiable protection against adversarial attacks (prompt injection, jailbreaking), unauthorized tool invocation, and data exfiltration.
  • EPI Evidence Mapping:
    • Ed25519 Cryptographic Signatures: Every .epi artifact's manifest is signed with Ed25519, ensuring the entire evidence package is completely tamper-evident.
    • SCITT Transparency Logging: Integration with SCITT logs registers the manifest's canonical SHA-256 hash in a public or private append-only transparency ledger, ensuring non-repudiation.
    • Tool Execution Capture: The steps.jsonl timeline captures every tool call, its parameters, and returned values in order, allowing auditors to verify that the agent never invoked unauthorized resources or executed unsafe commands.

2.2 Privacy and Data Governance (Control Domain 2)

  • AIUC-1 Requirement: Strict protection of sensitive data (PII, credentials) from being leaked, logged, or ingested for unauthorized training.
  • EPI Evidence Mapping:
    • Automatic Forensic Redaction: Built-in regex-based scanners in epi_core.redactor automatically scrub API keys, authorization headers, environment secrets, and PII from the execution steps before they are written to disk.
    • Data Boundary Isolation: The environmental context (environment.json) explicitly documents which runtime and package dependencies were used, verifying that training boundaries were respected.

2.3 Safety (Control Domain 3)

  • AIUC-1 Requirement: Prevention of out-of-scope, harmful, or unintended behaviors.
  • EPI Evidence Mapping:
    • Deterministic Step Chronology: steps.jsonl records all inputs, reasoning traces, and outputs in an index-sequenced, time-monotonic chain using prev_hash binding. If an agent drifts out-of-scope or behaves unsafely, the exact timeline is sealed and cannot be altered.

2.4 Reliability (Control Domain 4)

  • AIUC-1 Requirement: Consistency of performance and robust error handling.
  • EPI Evidence Mapping:
    • Error Continuation Auditing: EPI’s forensic analyzer seals both successful completions and raw exception traces, letting auditors verify how the system handled API failures, bad inputs, or rate limits.

2.5 Accountability and Transparency (Control Domain 5)

  • AIUC-1 Requirement: Human-in-the-loop (HITL) oversight, clear audit trails, and process transparency.
  • EPI Evidence Mapping:
    • Human Review Addendum: The review.json ledger provides a cryptographically bound log of human evaluations, sign-offs, and risk verdicts. This file is appended cleanly without modifying or compromising the original raw execution history.
    • Policy Preserving: The policy.json and policy_evaluation.json payloads travel with the container, preserving the exact rules and thresholds that evaluated the agent run.

2.6 Societal Impact (Control Domain 6)

  • AIUC-1 Requirement: Alignment of agent behavior with ethical boundaries and risk limits.
  • EPI Evidence Mapping:
    • Sealed Analyzer Findings: The analysis.json record captures heuristic and policy-grounded evaluations, providing a persistent, machine-readable proof of compliance for safety reviews.

3. Continuous Audit Verification

Using the epi verify command, an auditor can mathematically confirm an agent's continuous alignment with the AIUC-1 standard in seconds:

epi verify --policy strict loan-approval.epi

This command parses the container, recalculates all SHA-256 hashes, validates the Ed25519 signature, checks OTel sequence monotonicity, and confirms SCITT ledger registration—producing an objective, tamper-proof verification report that directly maps back to AIUC-1 control proofs.