Modern systems make claims constantly: this commit declared AI involvement, this model produced that output, this action earned that payout. Almost none of those claims come with evidence a third party can check. The most concrete case today: AI tools write an increasing share of production code, but git history can't answer which changes were AI-generated. Trailers are inconsistent, easy to strip, and not machine-verifiable.
AIIR (AI Integrity Receipts) is our first shipped product. It generates deterministic, content-addressed receipts for commits with declared AI involvement and verifies them anywhere: locally, in CI, or offline, without trusting a central service.
Public research extends the same receipt model further: inference receipts and receipted actions (attestable AI), plus evidence-first capsules in physics and mathematics, published Zenodo-first with open DOIs and explicit claim boundaries. See the Research page for the linked Zenodo records and reproducibility capsules.
| PyPI | pip install aiir |
| GitHub Action | invariant-systems-ai/aiir@v1 |
| GitLab CI | One include: line: push and MR receipts; GitLab Technology Partner |
| VS Code | Extension: local-first commit receipts, inline verification, receipt explorer |
| AI Assistants via MCP | Works with Claude, Copilot, Cursor, Continue, Cline, Windsurf |
| License | Apache 2.0; zero runtime dependencies (CLI core); Python 3.9+ |
Security posture: 2,499 collected tests, 100% coverage, and a public threat model with 150+ documented security controls, plus ClusterFuzzLite fuzzing, mutation testing, and conformance vectors. JSON and deterministic CBOR receipt formats. Optional Sigstore signing and PEP 740 attestations.
How it fits: AIIR fills the authorship-provenance gap before build-level tools (SLSA, in-toto, SCITT) kick in. See ecosystem positioning and the public Research page for the current public picture.
Detection is heuristic: AIIR records what is declared: Co-authored-by trailers, bot authors, and AI-tool markers. Agent-mode sessions such as Copilot Chat, Claude Code, and Cursor Agent do not add these markers today, so those commits can still land in the human bucket unless the tool or user declares them explicitly. See detection scope for current public limits.
Invariant Systems, Inc. is a Delaware C-Corp founded in 2025.
Website | Docs | Research | Browser Verifier | Contact: noah@invariantsystems.io
© 2025-2026 Invariant Systems, Inc.