Skip to content

luckyPipewrench/agent-evidence-levels

Repository files navigation

Agent Evidence Level (AEL)

A measurement standard for AI-agent audit evidence. AEL grades a record of AI-agent activity, from AEL-0 to AEL-4, by what an independent party can verify, and what omission they can detect, without trusting the vendor or the operator.

  • SPEC.md — the standard (CC BY 4.0).
  • docs/ARTIFACT-FORMAT.md — the artifact format the checker consumes.
  • docs/CHECKER-DESIGN.md — the check matrix and fixture matrix.
  • checker/ — the reference checker aelcheck and fixture generator aelgen (Apache-2.0).
  • fixtures/ — the per-rung conformance corpus: one valid artifact per rung plus the perturbed artifacts the checker must reject or flag (Apache-2.0).
  • GRADES.md — the self-grading registry; the editor's row is first and held to the same rule.
  • docs/VERSIONING.md — draft stability, compatibility, and donation posture.
  • docs/RATIONALE.md — non-normative positioning and adoption guidance, kept out of the spec.
  • docs/PRIOR-ART.md — how AEL relates to existing standards (transparency logs, audit-log and retention standards, evidence-handling), and what it deliberately does not restate.
  • docs/EXAMPLE.md — a worked end-to-end example: one real artifact from bytes to a grade.
  • schema/ — JSON Schemas for the record payload and manifest, for independent reimplementers.
  • CONTRIBUTING.md — contribution rules and required validation.

Why a checker ships with the spec

The standard's bar is "earned, not asserted": a grade counts only when a reference checker, run by someone other than the producer, demonstrates it against a real artifact, including rejecting a perturbed copy. A standard with that bar that shipped without a runnable checker would be the same attestation it criticizes. So the checker and the fixture corpus are part of v0.1, not a follow-up.

Run it

make build     # build aelcheck + aelgen
make gen       # regenerate the fixture corpus from a fixed test seed
make test      # go test ./...
make check     # regenerate + grade the whole corpus; assert every case matches its expect.json

make check printing, for each rung, the valid artifact graded at that rung and every perturbation rejected or flagged, is the proof the standard is earned.

aelcheck --keys <keysdir> <artifact>

grades one artifact and prints the full grade line plus a per-check PASS / FAIL / UNABLE-TO-VERIFY table.

Licensing

The specification (SPEC.md) is CC BY 4.0. The checker and fixtures are Apache-2.0. See LICENSING.md.

Status

v0.1 draft. This repository is vendor-neutral by construction: SPEC.md carries no product marks; concrete deployments (including the editor's) are graded in GRADES.md. The intended governance end state is donation to a neutral body after the draft has real independent use and critique.

See docs/VERSIONING.md for the draft stability, compatibility, and donation policy.

About

Agent Evidence Level (AEL): a vendor-neutral standard grading AI-agent audit evidence (AEL-0 through AEL-4) by what an independent party can verify, shipped with a runnable reference checker and conformance corpus.

Topics

Resources

License

Unknown, CC-BY-4.0 licenses found

Licenses found

Unknown
LICENSE
CC-BY-4.0
LICENSE-SPEC

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages