This page produces the public proof page's visual artifact honestly: it installs a real PocketAgent
persona via the production codec, boots a real LLM in this tab (WebLLM · WebGPU → Metal · no server,
no API key), asks a fixed question, and records the actual streamed reply as a clip. The text on screen
is the model's live output — not a script. Run it once, then drop the two downloads into
scripts/pocketagent/proof/.
scripts/pocketagent/proof/, then tell me — I'll run
node scripts/pocketagent/proof/run-behavioral.mjs --result scripts/pocketagent/proof/captured.json
and wire the clip into proof.html.