--- title: "Moonglade Provenance 2.5 is Open Source: Deterministic Verification for the Agentic Era" description: "We have open-sourced the reference implementation of Moonglade Provenance. Dual-language parity, RFC 6962 Merkle trees, RFC 3161 timestamps, and a sub-4ms Rust WebAssembly runtime—available now on GitHub." pubDate: 2026-10-19 author: "Anaiya" series: "The Agent Provenance & Safety Series" part: 6 readTime: "5 min read" heroImage: "/images/blog/provenance-open-source-release.jpg" image: "/images/blog/provenance-open-source-release.jpg" tags: - provenance - open-source - architecture - security - webassembly - agents audio: "/audio/2026-10-19-moonglade-provenance-is-open-source.mp3" --- ![An engineering workstation in St. Andrew, Barbados, displaying the public Moonglade Provenance repository, test suites, and hardware signing keys.](/images/blog/provenance-open-source-release.jpg) Today, we are making the reference implementation of **Moonglade Provenance** public on GitHub. You can inspect, clone, and build the release candidate directly: **[github.com/MoongladeAI/moonglade-provenance-public](https://github.com/MoongladeAI/moonglade-provenance-public)** This release marks the transition of our cryptographic attestation and drift-detection framework from an internal lab invariant into open, inspectable infrastructure. --- ## Why We Built This Every organization is deploying autonomous or agentic AI workflows. Few prioritize assurance about what their agents actually read, generated, or modified. When multi-agent systems operate inside live codebases, documentation repositories, and executive filesystems, unverified assertions compound invisibly. In our own studio vault, we watched autonomous agents cite documents that did not exist and place hardcoded "SEALED" strings over unverified directories. The industry’s default response has been stochastic evaluation: prompt engineering, secondary LLM judges, and probabilistic scoring. For an engineer with thirty years on the wire—running mail servers, authoring DNS zone files, and engineering enterprise tooling for C-suite decision surfaces—that response is unacceptable. You cannot insure an organization on stochastic optimism. When software acts, it must leave mechanical proof. Moonglade Provenance replaces guesswork with deterministic math: - If a document has not changed by a single byte, it is **0% Drift**. - If a signature was not computed over the exact canonical preimage, the gate **fails closed**. - If an agent asserts state without an RFC 6962 Merkle inclusion proof, the claim is rejected. --- ## What Is in the Repository The staging repository packages `@moonglade/provenance` (TypeScript / Node.js 22 & 24) alongside specifications and verification harnesses: 1. **Dual-Language Parity & Shared Golden Vectors**: The TypeScript and Python reference engines produce bit-for-bit identical RFC 6962 Merkle trees and canonical byte digests across shared test vectors. 2. **The Draft -01 Canonical Normalization Profile**: Text artifacts are strictly normalized (UTF-8 NFC, trailing newline normalization, CRLF to LF, and fatal handling of non-conforming encodings). Binary assets (images, ONNX weights, PDFs, compiled binaries) preserve exact byte fidelity. 3. **Ed25519 & OpenPGP Dual-Sealing**: Native support for fast Ed25519 signatures alongside traditional OpenPGP hardware-key workflows (YubiKey / GnuPG). 4. **RFC 3161 Hardware Timestamp Attestation**: Cryptographic timestamp tokens bind signed artifacts to trusted external time authorities, preventing back-dated state assertions. 5. **Sub-4ms In-Browser WebAssembly Verification**: The accompanying Rust micro-runtime (`mgc_wasm`, built on `ed25519-dalek`) brings real Edwards curve point decompression directly into client browsers, eliminating the need for server-side verification hops. 6. **Storage-Agnostic Tiers & Database In-Memory Verification**: Provenance applies to any sequence of bytes and lives anywhere you keep JSON: as a detached .provenance.json sidecar (Tier 1), aggregated into an RFC 6962 Directory Registry (Tier 2), or stored directly inside an in-memory database row (Postgres JSONB, Redis KV, or API payload, Tier 3) with zero disk requirement. Eliminates 1:1 sidecar file clutter across large workspaces by aggregating thousands of files into a single root registry with $O(\log n)$ verifiable inclusion proofs. --- ## What a Cryptographic Seal Establishes (And What It Does Not) In keeping with our core epistemic commitments, we state plainly what this tool does and does not do: - **What it establishes**: A valid signature proves that the exact canonical bytes of an artifact were witnessed and attested by the holder of a specific cryptographic key at a verifiable point in time. It guarantees tamper-evidence and bit-level integrity under that key. - **What it does not establish**: A cryptographic signature cannot tell you whether the text in a file is true, whether the author was competent, or whether an AI model performed better because it used provenance. Those require human domain judgment and controlled scientific evaluation. We do not sell magic. We provide the mathematical ruler so that you can verify the boundaries yourself. --- ## Getting Started To run the test suite and verify the package locally: ```bash git clone https://github.com/MoongladeAI/moonglade-provenance-public.git cd moonglade-provenance-public npm ci npm test npm run self-test ``` The test harness executes 118 unit, property, and integration tests, validating Golden Vector parity, RFC 6962 tree calculations, and tamper rejection. --- ## A Sovereign Baseline We engineered this framework in St. Andrew, Barbados, not to adapt foreign hand-me-downs, but to author a hardened foundation capable of commanding international respect. Our commitment to regional and global engineering is simple: **software that proves its work, built by practitioners who know what fails under pressure.** Inspect the code, run the tests, and verify the math: **[MoongladeAI / moonglade-provenance-public](https://github.com/MoongladeAI/moonglade-provenance-public)**