The Bench
Interactive experiments you can run in the browser. Read the Field Note, then come break it here.
Loop Lab
The companion to FN-004. A loop engineering agent that wants to "clean up" a database. Toggle the three external fences and watch the same agent either delete production or get stopped. The brake is not in the model; it is the layer you switch on here.
Toggle the fences and run. With the gate off, the hallucinated-done agent deletes the database. With it on, the same agent cannot, by construction.
Deterministic, no API keys, same logic as the Python tests. Source: src/fn04.
Pick a sample query and replay a real pre-recorded agent trace, token by token: tool calls, reasoning steps, and the final response. No live model calls.
More on the Bench
Every Field Note and Lab Note with an interactive piece lands here. Next up: the eval-rubric failure-bucket explorer from FN-002, and the multi-agent versus router comparison from SG-003 and the Lab Note.