◈ Assistant Behavior Checks
Confirms Aperio's core conversation loop still behaves correctly after every code change — using a scripted, fake conversation instead of a real AI model, so every check is exact and gives the same result every time it runs.
How to read this page: each check below plays out one specific real-world situation — finishing a normal multi-step task, catching itself after a false claim, giving up after repeated errors, handling a huge result safely, refusing to act on untrusted content, and breaking out of a failure loop. A red check means a code change broke one of these behaviors, not that an AI model made a bad judgment call.
This is the fast, repeatable sibling of the live model-tier benchmarking exam: that page answers "is this AI model good enough to drive Aperio?"; this page answers "did today's code change break how the assistant behaves, regardless of which AI model is used?".
tests/harness/ · npm run test:harness