Judge Quick Start

Three tests. Both sides of the WebMCP contract.

Nothing to author. Run a healthy baseline, a disclosed site-side fault, and a real semantic collision—then inspect one complete Judge Results report.

Build here in any browser. Run the copied command in a fresh GPT-5.6 Sol or Terra chat using ChatGPT Desktop's built-in Browser—not the Chrome side panel.

Preloaded three-case contract

Know what each test is proving before you run it.

Test 1 · Live agent baseline

Expected pass

“Set the Stoneware mug quantity to 3.”

Normal handler. Expect cart_update, mug quantity 2 → 3, revision 0 → 1, and one ledger transition.

Test 2 · Controlled planted site fault

Deterministic issue

“Set the Field notebook quantity to 2.”

The session-only demo handler deliberately reuses the current quantity and returns a successful no-op. Thurstone should catch the missing required effect.

Test 3 · Live agent semantic stress test

Outcome not predetermined

“Show me my current order.”

cart_get deliberately overlaps with order_review. The real agent may preserve or miss the owner's intended boundary; Thurstone records what actually happens.

You see the expected behavior here. The fresh agent receives only each request and the live catalog—not the owner's answer key.