AN EXPORT, ALLEGEDLY.
Done
or not?
The API said success.
What did that actually prove?
Six document-export traces. Only one satisfies the full task contract. Choose a trace, inspect the evidence, then reveal the verdict.
THE CONTRACT Produce report-final.pdf and confirm it can be read.
CHOOSE A TRACE
Synthetic examples.
No account, keys or API calls.
TRANSPORT RECEIPT
≠
COMPLETED
TAKE IT INTO YOUR WORKFLOW
Write the completion rule
before the success message.
A transport response, a tool result, a job state and a usable artifact answer different questions. Record the one your task actually depends on.
node bin/check.mjs examples/completed.jsonThe same checker powers this page and the CLI. Read the contract, change a fixture, run the tests.
Explore the code →What this experiment proves
These fixtures illustrate a narrow export contract. The checker classifies supplied evidence; it does not contact a server, verify a real file, judge document quality or establish that a trace is truthful. It is not an agent benchmark or an MCP conformance test.
Protocol example: MCP distinguishes tool execution errors from protocol errors. See the official SDK documentation. The four-layer contract here is specific to this experiment.