Adversarial agent evaluation

BoundedOps

SYNTHETIC OFFLINE — ZERO PROVIDER REQUESTS, TARGET EXECUTIONS, AND REAL SIDE EFFECTS

Turn declared mission and operating boundaries into 24 fixed adversarial tests, deterministic findings, review-gated remediation, same-seed regression, and a version-bound development attestation.

Runtime and source boundary

Declared Git metadata is shown separately from the measured runtime-source digest. Exported source does not claim cryptographic Git proof.

Loading bounded runtime identity…

1. Start assessment

Starting runs only the vulnerable synthetic actor through the no-side-effect gateway. It stops at the human gate.

Ready. No assessment has run in this browser session.

5. Verify and download evidence

Verification rechecks the development signature, immutable restricted-evidence references, and material-change staleness. The download is a separately privacy-validated public JSON envelope.

Download public evidence

Claim boundary

This local/container-ready interface is not proof of Cloud Run, successful Gemini, authenticated ADK execution, live Firestore or Cloud Storage persistence, managed observability, production signing, public judge access, or submission.

BoundedOps

Runtime and source boundary

1. Start assessment

2. Inspect results

Mock gateway attempts

Findings

Decision and condition history

Signed subject binding

3. Record an attributable decision

4. Scoped resume

5. Verify and download evidence

Claim boundary