openveritas
Verify what the agent actually did — not what the screenshot shows.
A ground-truth state oracle for computer-use agents via the OS layer (a11y tree + filesystem + network).
You can't safely run an agent on real software if you can't verify what it actually did. Screenshot-based judges miss non-visual side effects — file metadata, DB rows, network calls, an accidentally-followed account — and pass agents that reached the right pixels via the wrong path.
Invert verification: instead of judging screenshots, instrument the OS (accessibility tree + filesystem + network) to emit deterministic before/after “action receipts” as ground truth — plus a content-addressable semantic action codec that makes GUI actions portable and version-robust.
OS-layer instrumentation
Captures structural state from the a11y tree, filesystem, and network — below the pixels.
Action receipts
Deterministic before/after state diffs serve as ground-truth verification of each step.
Semantic action codec
Content-addressable action encoding makes GUI actions diffable, portable, and resolution/version-robust.
Eval & audit
A CI gate for computer-use agents and a tamper-evident audit trail of real side effects.