B
BOUNDARYDeterministic agent control
Simulated tools only

OPENAI BUILD WEEK · JUDGE WORKSPACE

AI that asks
before it acts.

Interpret policy with GPT-5.6 or a synthetic fixture. Confirm it as a human. Then watch deterministic code redact, route, pause, block, and audit every proposed action.

ReadyStart by interpreting the synthetic policy.

Ephemeral demo state. Sessions reset after 30 minutes of inactivity or any deployment restart.

01

Author policy

NO API KEY
288 characters
02

Review proposal

WAITING
Interpret the authored policy to create a reviewable, non-authoritative draft.
03

Human confirmation

REQUIRED

GPT-5.6 can propose rules. It cannot activate them. A separate reviewer establishes human provenance.

04

Compiled policy

INACTIVE
No policy is active. Unconfirmed interpretation cannot evaluate or execute actions.
05

Action simulator

SYNTHETIC INPUTS

LATEST DECISION

No action evaluated yet

Select a preset after confirmation.

06

Approval console

CLEAR
Actions above policy thresholds will pause here for a human decision.
07

Adversarial analysis

NON-EXECUTING

Reviewed bypass suggestions become deterministic fixtures. They never invoke tools.

08

Append-only audit timeline

0 EVENTS
Compilation, evaluation, transformation, approval, and simulated execution events appear here in order.