AgentBench / ClawCheck · Overview

Status — audited-prototype reality

This page describes design intent for the DefendableOS ecosystem. In the current independently audited prototype this capability is ROADMAP / NOT YET IMPLEMENTED as a public/production service unless explicitly shown as fielded. The only fielded, Codex-audited surface today is the controlled synthetic agent-operations demo — see Field Proof v0.1. NOT CLEARED FOR PRODUCTION · NOT CLEARED FOR EXTERNAL SAAS ENFORCEMENT. SHA-256 hashes here provide content-integrity linkage only — not signatures, owner approval, external attestation, or blockchain anchoring.

AgentBench and ClawCheck are the planned agent-testing surface of the DefendableOS ecosystem. They are design intent, not a fielded service.

Roadmap

The capability described below is not built or deployed. It is the intended shape of the surface, documented honestly so the books and records show the plan, not a claim. The live, hash-chained proof surface today is DefendableCloud — the Defendable Run and its eval flight-sheet lane at api.defendablecloud.com.

AgentBench is designed to let you submit an agent, run it against a standard benchmark pack, and return a structured verdict. The verdict is planned to integrate with the Tribunal roadmap (a declared rulebook applying flags, not a judge model) — it will be a ruleset-audit outcome, never a quality opinion. ClawCheck is designed as the adversarial-security companion: a set of CVE-sourced agent-security attack templates and a structured defense pattern for exercising an agent under hostile inputs.

Until these land, the built analogue is the DefendableCloud verdict tiers (honey = pass · jelly = risk · propolis = fail) and per-org hash-chained receipts.

---

🐝 Operator-grade · books and records · to the shed.