2026-05-17 22:27 UTCIn-site rewrite4 min readUpdated: 2026-06-30 13:03 UTC

EPI – forensic evidence containers for AI agent SCITT-compatible, EU AI Act-ready

EPI packages AI agent execution into a cryptographically signed, tamper-evident forensic container (.epi file) for regulatory compliance and offline verification. It integrates with popular AI frameworks and maps to EU AI Act, FDA 21 CFR Part 11, SOC 2, and more.

SourceHacker News AIAuthor: afridi_epilabs

Notifications You must be signed in to change notification settings

Fork 4

Star 24

BranchesTags

Open more actions menu

Folders and files

NameName

Last commit message

Last commit date

Latest commit

History

348 Commits

.github

config

demo_workflows

docs

epi-viewer

epi_analyzer

epi_cli

epi_core

epi_gateway

epi_guardrails

epi_recorder

epi_viewer_static

examples

pytest_epi

scripts

simulation

tests

web_viewer

.aiignore

.env.example

.gitignore

CHANGELOG.md

CLAUDE.md

CONTRIBUTING.md

CONTRIBUTORS.md

Caddyfile

Dockerfile

LICENSE

MANIFEST.in

README.md

SECURITY.md

conftest.py

docker-compose.prod.yml

docker-compose.yml

epi.spec

epi_postinstall.py

pyproject.toml

setup.py

technical_overview.md

Repository files navigation

Portable Evidence for AI Execution.

Pilot Program · Quick Start · Integrations · Regulatory Mapping · Standards · Docs

When a regulator asks what your AI agent did six months ago,

the answer should be a file — not a shrug.

🏛️ The Problem

AI agents make decisions that carry legal, financial, and safety consequences. Those decisions happen in memory, get logged in transient cloud infrastructure, and are tied to a runtime that may not exist when an auditor arrives six months later.

EU AI Act Article 12 requires providers of high-risk AI systems to maintain logs of operation appropriate to the system's lifecycle. FDA 21 CFR Part 11 requires tamper-evident audit trails. SOC 2 CC7.2 requires logging of unauthorized activity. None of these regulations define how evidence must be packaged for external review.

EPI closes that packaging gap.

💎 What EPI Does

EPI packages AI agent execution—the complete decision trail, governance evaluation, tool calls, inputs, outputs, approvals, and environmental context—into a single .epi artifact: a cryptographically signed, tamper-evident, self-contained forensic container.

The artifact can be emailed to an auditor, archived for 10 years, or opened on an air-gapped machine—without calling home, without the original runtime, and without trusting the producer.

The Forensic Container Architecture

agent_run.epi ├── [Envelope] EPI1 Header — Binary magic, version, and payload SHA-256 └── [Payload] Signed ZIP — Wrap with EPI1 Envelope ├── manifest.json — Ed25519 signed root of trust + file hashes ├── steps.jsonl — Immutable execution timeline (prev_hash chain) ├── governance.json — The rulebook that governed the run ├── environment.json — Host & Python runtime context snapshot ├── artifacts/ — Raw evidence preserved verbatim (e.g., AGT bundles) ├── viewer.html — Self-contained offline forensic viewer └── VERIFY.txt — Human-readable offline verification guide

🚀 Quick Start

pip install epi-recorder

Instrument an Agent

Capture any LLM-backed workflow with minimal boilerplate.

from epi_recorder import record, wrap_openai from openai import OpenAI

1. Wrap your client for high-fidelity capture

client = wrap_openai(OpenAI())

2. Record the forensic context

with record("loan-approval.epi", workflow_name="Credit-V4"): response = client.chat.completions.create( model="gpt-4o", messages=[{"role": "user", "content": "Assess applicant #421"}] )

Verify the Artifact

epi verify --strict loan-approval.epi

──────────────────────────────────────────────────── EPI Verification Report (v4.1.0) ──────────────────────────────────────────────────── Trust Level: HIGH Signature: VALID (Ed25519) Integrity: OK (SHA-256 manifest, 23 files) Chain: INTACT (prev_hash verified, 47 steps) Identity: KNOWN (production-signer-v4)

This artifact has not been modified since sealing. ────────────────────────────────────────────────────

Open the Offline Viewer

epi view loan-approval.epi

Opens viewer.html in your browser — zero dependencies, offline-ready.

🔌 Integrations

EPI acts as a universal evidence layer for the entire AI ecosystem:

LangChain: ChatOpenAI(..., callbacks=[EPICallbackHandler()])

LiteLLM: litellm.callbacks = [EPICallback()]

Microsoft AGT: epi import agt — adapter for regulatory evidence.

OpenTelemetry: setup_epi_tracing() — turn spans into signed artifacts.

pytest: pytest --epi — automatic forensic evidence for failing tests.

⚖️ Regulatory Compliance Mapping

EPI produces evidence that addresses specific global regulatory requirements. EPI is not a compliance guarantee and does not provide legal advice. Whether the enclosed evidence satisfies a specific regulatory threshold is for the auditor or notified body to determine.

Requirement Framework .epi Evidence

Logs of operation appropriate to lifecycle EU AI Act Art. 12 steps.jsonl + environment.json

Technical documentation retention (10yr) EU AI Act Art. 19 Sealed .epi (format-stable)

Evidence of Human Oversight EU AI Act Art. 14 review.json approval ledger

Audit trail for regulated software FDA 21 CFR Part 11 Signed steps.jsonl + manifest.json

Non-repudiation of data HIPAA § 164.312 Ed25519 signature over manifest

AI Risk Management documentation NIST AI RMF governance.json + analysis.json

Verifiable risk evaluation & HITL audit proof AIUC-1 (SOC 2 for Agents) steps.jsonl + review.json + analysis.json

🏛️ Standards Alignment

SCITT (IETF): EPI produces SCITT-compatible COSE Sign1 statements for transparency log anchoring.

AIUC-1 (Compliance): EPI generates structured compliance audit evidence mapping to the six AIUC-1 risk domains for autonomous agent audits. See AIUC-1 Evidence Mapping for details.

CycloneDX: Preserves CycloneDX SBOMs under artifacts/sbom/ for software supply chain transparency.

in-toto (CNCF): Roadmap: Exporting steps.jsonl as in-toto link files for execution supply chain verification.

Ed25519 (RFC 8032): All manifests are signed using industry-standard Ed25519 digital signatures.

🤝 Founding Pilot Program

EPI is seeking regulated enterprises to pilot AI compliance evidence packaging.

If you operate AI agents under the EU AI Act, FDA 21 CFR Part 11, or SOC 2, and you need portable, independently-verifiable evidence, we invite you to join our Pilot Program.

What Pilots Receive:

Direct Integration Support: Hands-on assistance from the maintainers.

Priority Roadmap Influence: Shape the standard based on your compliance needs.

Founding Partner Recognition: Optional listing as an early adopter.

Contact: [email protected] — Subject: EPI Pilot — [Your Organization]

📑 Documentation

📖 Protocol Specification: The technical wire format.

⚖️ Governance Guide: Managing rulebooks and evaluations.

🇪🇺 EU AI Act Prep: Evidence workflow guide.

🛠️ CLI Reference: Comprehensive command guide.

🛡️ Security Model

Threat Mitigation

Post-Seal Tampering SHA-256 file manifest + Ed25519 signature.

Evidence Replay Unique workflow_id + time-anchored created_at.

Secret Leakage Automatic forensic redaction of API keys, tokens, and PII.

Signature Spoofing Strict ed25519:: format enforcement.

Built by EPI Labs.

Ensuring that as AI moves faster, accountability stays ahead.

MIT License | Contributing | Security Policy

About

EPI (Evidence Packaged Infrastructure) packages AI execution as evidence.

www.epilabs.org/

Topics

epi

ai-agents

ai-systems

aisystem

ai-evidence

evidence-packaged-infrastructure

Resources

Readme

License

MIT license

Contributing

Security policy

Uh oh!

There was an error while loading. Please reload this page.

Activity

Stars

24 stars

Watchers

0 watching

Forks

4 forks

Report repository

Releases 3

v4.1.0 — SCITT Transparency Service Integration

Latest

May 15, 2026

+ 2 releases

Packages 0

Uh oh!

There was an error while loading. Please reload this page.

Contributors

Uh oh!

There was an error while loading. Please reload this page.

Languages

Python 89.4%

JavaScript 5.8%

PowerShell 2.2%

CSS 1.7%

HTML 0.8%

Shell 0.1%