Operator console

From evidence to accountable action.

A proposal is evaluated, authorized, executed, verified, and recorded without giving the agent approval authority.

Synthetic sampleNo content leaves this browser.

Synthetic sandbox · not production authorization

Evidence
Proposed action
Policy verdict
Approval
Receipt
Rollback

Current proposal

Evidence

Target

Deterministic policy verdict

Rule
Adapter
Environment

Prepared review

Engineering proof

Tool access is not authority.

Closed catalogNo shell strings or model classification

Exact approvalRequest + decision bound, expiring, single-use

35 adversarial casesStructured expected outputs

Verifiable receiptsStrict schema + digest; synthetic adapter labeled

What this proves: deterministic policy can veto or gate typed synthetic actions and leave inspectable evidence.

What it does not: provide production authorization, tamper-proof audit logs, machine isolation, or truth guarantees.

Operational proof: a private scheduled shadow path passed 10/10 supervised cases. One green follow-on now reads a hash-bound local evidence file through a confined read-only adapter and writes one immutable receipt. The live receipt verified with no approval, matching source and before/after hashes, and an idempotent replay changed no files. Yellow execution remains disabled.

Portable green proof: download the frozen v2 proof packet generated by the actual review and synthetic executor path. Its strict schemas and recomputed digests show internal integrity and deterministic policy binding—not authenticity, production authorization, or an external action.

Human boundary proof: a separate synthetic run required a literal interactive approval after presenting the exact effect and rollback contract. The single-use grant produced one verified sandbox receipt; replay returned the same receipt without a second effect. No external system was connected.

Evidence before authority: inspect the upstream reasoning in the Context Layer Lab.