AegisProof

Open, vendor-neutral evidence packages for AI systems.

AegisProof is a small specification and CLI so teams can declare what evidence exists for an AI system revision (aegisproof.yaml), hash those artefacts, and verify integrity in CI — without an LLM, without a SaaS, and without claiming legal compliance.

Integrity ≠ compliance. A green verify means listed files match their digests. It does not mean the system is safe, certified, EU AI Act compliant, or “assured.”

If you ship LLM/RAG/agent features into enterprises, you already assemble evals, policies, test outputs, model cards, and SBOMs for security questionnaires. AegisProof turns that pile into a portable, diffable package other tools (and humans) can check.

---

Why fork / contribute

• Standards-shaped OSS — schema + conformance + boring CLI (Apache-2.0)
• Compose, don’t reinvent — point at SPDX / CycloneDX / SARIF; don’t replace them
• Local-first — works offline; cloud is optional and not required
• Clear non-goals — not a GRC dashboard, not RiskLine, not “AI SBOM”

Good first moves: run the examples, add a conformance fixture, emit a package from your AI app’s CI, open an issue.

Docs / wiki: docs/wiki/Home.md · Issues: open issues · Spec: SPEC.md

---

Quick start

Requires Go 1.22+.

git clone https://github.com/new-world-coder/aegisproof.git
cd aegisproof

go test ./...
go run ./cmd/aegisproof verify ./examples/minimal
go run ./cmd/aegisproof verify ./examples/minimal --format json

Scaffold your own package:

go run ./cmd/aegisproof init ./my-system
# replace evidence files with real artefacts…
go run ./cmd/aegisproof pack ./my-system
go run ./cmd/aegisproof verify ./my-system

Minimal manifest shape

specVersion: "0.1"
system:
  name: my-ai-system
  version: "1.0.0"
producer:
  name: my-team
createdAt: "2026-09-16T12:00:00Z"
evidence:
  - id: eval-results
    type: evaluation
    locator: evidence/eval.json
    digest: "sha256:…"   # filled by `aegisproof pack`

---

What this is / is not

Is	Is not
Evidence package + digest verify	Legal compliance engine
CI-friendly interchange (SARIF-like role)	“AI SBOM” (use SPDX/CycloneDX for inventory)
Regulation-neutral core	EU AI Act / NIST encoded in core schema
Human-readable YAML/JSON	LLM-required tooling
Optional consumer for tools like RiskLine	A RiskLine feature / proprietary format

---

Repository map

Path	Purpose
SPEC.md	v0.1 specification
schema/	JSON Schema
cmd/aegisproof	CLI (verify, pack, init)
examples/	minimal, RAG, agent packages
conformance/	valid / invalid fixtures
docs/	architecture, research, roadmap, security

Exit codes

Code	Result
0	VALID
1	INVALID
2	STALE (digests OK; freshness warnings)
3	Usage / IO error

---

Related projects (not dependencies)

• RiskLine — deterministic AI risk classification / assure (may consume AegisProof later; does not own this standard)
• SPDX / CycloneDX — inventory & attestations to reference, not replace
• XERJ — optional retrieval/indexing experiments only

---

Status

Early v0.1 draft. APIs and schema may change before 1.0. Feedback via Issues is welcome — especially emitter/consumer designs and conformance gaps.

Planning background: docs/IMPLEMENTATION_AND_TESTING_STRATEGY.md.

Licence

Apache-2.0 — see LICENSE.