Recomputable Evidence
Section 1 — Mission and Scope
Lab Name
Short Description
A shared test bench for checking whether separate specifications for AI agent evidence still hold up when they are used together.
Scope of Lab
The problem. When an AI agent takes an action — makes a payment, calls a tool, acts for someone — several groups are now writing specifications for the record it leaves behind: who authorized it, what was done, who holds the keys. Each spec can be verified on its own. Real systems combine them: one spec records the authorization, another the payment, a third the custody of the keys. Few combinations have been checked, and each check was set up by one project under its own criteria. The typical failure is quiet: one record accepts another’s claim of authority as if it were its own, and the result looks valid when it isn’t. There is no shared set of criteria for judging a combination that the specs’ maintainers wrote together.
What the lab does. It keeps two things in one neutral place:
- Test cases that pair records from two different specs and state the expected outcome, including cases that must fail. Anyone can rerun them offline from the published bytes: no hosted service, no account.
- Criteria for judging a combination, written jointly by the people who maintain the specs involved, not by any single author.
What it does not do. It does not pick a winning spec, does not certify anyone, and does not take over any spec. Each spec keeps its own repository and maintainers.
Alignment with LFDT Mission
LFDT builds neutral, jointly governed infrastructure. This lab applies that to testing: a place where competing specs are checked against each other under criteria none of them owns.
Relation to Existing LFDT Labs and Projects
Agent Authority Conformance is the closest approved lab. It already hosts cross-stack test families, including one for action-ref and a run of a composition pack. What it does not hold yet is criteria for judging a combination, written jointly by the maintainers of the specs being combined; its vocabulary stays with its maintainer. We propose to collaborate: this lab references test families and run records hosted there by pin instead of copying them, and uses the same record rules (author-produced vs. independent runs, records never removed, a merge is not a verdict), so a record means the same thing in both places. We opened Agent-Authority-Conformance/aps-conformance-suite#121 to discuss it; nothing is agreed yet.
Proof-of-Control (verifiable evidence of AI agent behavior) and Proof-of-Process (attestation of how a digital action was produced) work nearby. We would like to explore shared test cases with both once this lab is set up.
Section 2 — Lab Details
Does this lab produce code?
Yes: test cases and a small verifier to run them.
Does this lab produce a specification?
Not initially. The criteria ship as text alongside the test cases. The specs being tested keep their own licenses and governance; if the criteria later grow into standards work, we will raise that with the stewards as a separate step.
Pre-existing Repository
https://github.com/giskard09/action-ref-conformance (Apache-2.0). Most of its commits have no DCO sign-off; we will squash the history into one signed-off commit when the lab repository is created.
Initial Committers
- https://github.com/giskard09
- https://github.com/azender1 (SafeAgent — our integrator partner under a signed Revenue Share Agreement. He recuses from any decision that would directly benefit his own commercial interest.)
- https://github.com/magentixai (Martin Sansone — CEO, Magentix.AI; author of the Enhanced Fraud Data standard, Pay.UK; also steering the x402 Foundation TSC evidence-record charter, a related but separate track)
- https://github.com/kenneives (AgentGraph/AgentAvow; wrote the independence rubric)
Licensing
☑ I understand that all code hosted in LFDT Labs must be made available under an Apache 2.0 license with DCO sign-off. ☑ I understand that all specification and standards work in LFDT Labs is done under the Community Specification License 1.0 and the rest of the CSL framework.
Governance Model or Practice
Written down in GOVERNANCE.md. The core rule: no criterion or test case is merged until a committer other than its author has rerun it from scratch. We already work this way: one committer wrote the first rubric, a second reran it on a fresh clone, and a third found a non-neutral detail that was fixed before merge.
Security
No formal vulnerability process yet; we will add one before the first release.
Infrastructure and Tooling
GitHub Actions CI on the seed repository. No mailing list or chat yet.
Evidence of Adoption and Use Cases
- SafeAgent (azender1, a founding committer) records
action_refevidence in production: a live deployment, not a demo. - The action-ref test family hosted in the Agent Authority Conformance lab was rerun by an outside runner who authored neither the vectors nor the implementation, recorded there as independent (commit a7aa94f).
- The JSON canonicalization fixtures one committer published (magentixai/axes,
portable/jcs-properties) were checked against an independent library,rfc87850.1.4: 4 of 4 agree.
Roadmap
Next 6 months: settle the collaboration with Agent Authority Conformance; publish the first paired test case between two separately maintained specs; add a vulnerability process.