Live server evidence, not a test or replay provider.

Each side ran the checkout with T3CODE_DEV_SEED=0 vp run dev. State was created through MCP; no live user database was copied. A real authenticated Claude SDK query generated 200 arithmetic lines. While its command lifecycle was started and the thread was running, the MCP driver changed effort from high to xhigh and observed through the provider result.

The MP4 files render the captured terminal transcript at its original timestamps; they are not browser recordings. session.cast is the source terminal recording. The protocol excerpts retain only relevant actual native events; model identifiers and unrelated payloads are omitted, and provider session IDs are consistently aliased within each side.

Driver usage: node driver.mjs <checkout> prove <new-output-directory>
The sanitized driver selects the last advertised Claude model (the same catalog entry used during these runs) without embedding its identifier. Requires dependencies installed and the checkout dev server running against its own .t3 state.

This proves successful delivery to the existing running SDK query. It does not measure thinking quality, prompt-cache reuse, or later-turn behavior.

AI-assisted; verified with the tests and tools listed above.