eval_loop

Overview ¶

Example: eval_loop demonstrates the eval-driven development loop (WC-C-001): capture a dataset from a real run, run it against an agent target, score it, and gate on regressions — the trace → dataset → eval → gate cycle.

It runs with NO API keys and NO network: a stored session is seeded in an in-memory SQLite store, a dataset is captured from it, and two deterministic "agents" are scored — a good one that passes the gate and a regressed one that the gate catches (the same non-zero signal `chronos evals gate` gives CI).

go run ./examples/eval_loop/