AER-1 interop
A standard is real when other people build it. Six did.
AER-1, the agent execution receipt format published as draft-zambo-aer1-26 at the IETF, was written to be implemented by strangers. This is the public record of the strangers: independent implementations built from the draft text alone, each checked against the conformance corpus, plus the verifications and the cross-system run. Every entry below is in the public registry, aer-1/implementations.json in the kit repo.
Conformant independent implementations
| Implementer | Language | Verified result | Against |
|---|---|---|---|
| ARION | Node.js, stdlib only | 165 of 165 vector checks, 0 failures; written from the draft text alone; three findings adopted into draft -09 | draft -08 |
| ARION | Java 11+ | 186 of 186 vectors; the first -10 conformance recorded anywhere | draft -10 |
| ColonistOne (aer1-perl) | Perl, core modules only | 68 of 68 vectors, zero disagreement with the reference runners; three findings adopted into -10 and credited in its acknowledgments | draft -09 |
| Genesis | Python, stdlib only | Built cold from the text: 9 of 9 valid and 28 of 28 invalid cases scored correctly; three spec gaps filed, all adopted into -20 and credited there | draft -19 |
| WhiteMagic Labs | Python / Rust | Independent fixture probe, six checks, all pass, digest recomputed byte for byte; validator converged with the reference kit across the full case matrix; listed as independent validator in draft -22 | draft -20 to -22 |
| Laurent Lemonnier | Python | 119 of 119 checks across the seven base vector files against kit v1.5.7, from the -26 text alone; re-validated October 31, 2026, still 119 of 119 | draft -26 |
Parity-checked, and labeled as exactly that
Three more ARION ports were run against the reference implementation for parity, which is a weaker claim than kit conformance, and the registry says so: Ruby (24,834 parity comparisons, 0 divergences), Swift (19,615, 0 divergences), and PHP (11,529 runs, 0 verdict divergences). A fourth entry, an anchor-witness in Python, is registered with no counts recorded. Parity is not conformance, and this table does not blur the two.
Independent verifications
Matthew Moore cross-tested draft -07 with the VLC-1 work: all 14 Section 7 attack cases produced exactly what the draft calls for. Cairn Commons recomputed draft -14 from scratch, their own script, hashlib only, no network: the workflow Merkle root and the example output hash matched, and three independent repliers reproduced it again. Verification that starts from zero trust in the author is the kind that counts.
Cross-system interop
On October 8 and 9, 2026, an AER-1 receipt chain was verified end to end against Tanilo's system: 15 of 15 aer1-verify checks, the Tanilo canonical hash independently recomputed and matching, its signature verifying against Tanilo's live keys, anchored on GOAT mainnet block 15948054. Two receipt systems, built separately, checking each other's work.
The honest gap, and the invitation
No independent implementer has a recorded run against the full current corpus yet: 143 vectors, 146 checks at kit v1.5.7. Laurent's 119 of 119 covers the seven base files. Running the rest is now a one-line job: the kit is a GitHub Action.
uses: zambodotdev/aer1-kit@v1 with mode: corpus runs every vector in your CI; mode: cross runs the agreement matrix; mode: receipt verifies a single receipt offline. Build the seventh implementation, run it, and the registry takes the entry. The corpus lives at zambo.dev/aer1/corpus.json, and the production system behind it, 153 tools behind one MCP endpoint with a verifiable receipt for every job, is at zambo.dev/add.