Native adapter receipt
No operator-triggered executeOnce call yet. Automatic compatibility calibration is shown in the readiness receipt and is not model-selection evidence.
Live WebMCP path
Run one sealed decision through the page's real tool catalog.
Open this route in the official ChatGPT in-app browser. In Chrome 149+, open chrome://flags/#enable-webmcp-testing, choose Enabled, and relaunch Chrome.
If Thurstone reports consumer-mismatch, close other same-origin Thurstone tabs, then reload this tab.
Declared fixture
checkout-seed-v1 · r0Seeded checkout sandbox
- Field notebookfield-notebookCurrent × 1$18.00
- Stoneware mugstoneware-mugCurrent × 2$48.00
Subtotal · standard shipping $7.00$73.00
UI receiptNo UI operation yet.
Runtime truth
WebMCP capability matrix
- Secure context
- Required
- Site Tools provider · registerTool()
- Unavailable
- In-page discovery · getTools()
- Unavailable
- In-page execution · executeTool()
- Unavailable
- Direct ChatGPT path
- Unavailable
- Chrome model-selection path
- Unavailable in this document
RegistryidleNo verified native tools detected.Generation 0
Gate 7 · no-key model-backed judge lane
Checking allocation…One fixed decision, one verified native read
Open this page in a supported Chrome/WebMCP consumer and wait for consumer-ready native discovery.
checkingChecking the durable judge allocation…
Judge diagnostics
The server accepts no prompt, model, schema, URL, or tool choice from this page. Its only request is: Which current cart lines have a quantity greater than one?
This is a separate, non-scored demonstration. A fresh decision can consume the single challenge-lifetime judge allocation. Afterward, the sealed decision remains available for local native replay without another model call.
Judge lanecheckingChecking the durable judge allocation…Judge status is not available.No sealed judge decision is exposed.
Lab plumbing, reset receipts, and Gate 1 proof
Gate 1 plumbing only · not model selection
Direct native WebMCP controls
Cancellation is recorded at both boundaries. Chrome may reject the consumer call after the harmless cart_get handler has already completed; the journal preserves that completed no-effect handler trace separately and never relabels it as handler cancellation.
Latest native handler trace
No native handler trace yet.
Append-only domain evidence
0 eventsLatest operation and reset receipt
Latest canonical operation trace
No operation trace yet.
Reset verification receipt
No hard reset receipt yet.
UI and native handlers share one serialized checkout store. Direct expected calls above are plumbing evidence only and are never counted as semantic model-selection evidence.
One document · one attachment
Gate 1 proof bundle
One click reloads into a clean document, waits for every exact runtime boundary, runs the fixed ten-call native sequence and one verified reset, records step timestamps and durations, validates the strict Gate 1 completion predicate, and requests one JSON download. It includes no cookies, browser history, account data, or automatic upload.
Ready to start a clean one-button proof run. Updated not started
This is native-plumbing evidence only. Its hashes prove internal consistency, not external attestation, model selection, semantic scoring, or Direct ChatGPT behavior. The automatic download is requested locally; the verified in-memory bytes remain available for retry if the browser blocks it.
Gate 2 · model selection
InactiveFinal fresh-context Probe calibration
The bounded lane runs four explicitly non-scored synthetic requests. Each uses a fresh Lab document, one stateless model decision, at most one native WebMCP call, verified resets, and sealed evidence.
Guard status loads when this deep calibration control enters view.