OptionalexperimentPrefix for the LangSmith experiment name.
OptionalextraExtra scorers appended to the built-in status_match + no_fabricated_documents — the caller
threads in the runner-agnostic evaluators (answer_correctness, true_pass) so a V2 experiment
is scored by the same metrics as V3. Kept generic (no @repo/ai-agents dep here); the CLI wires
them together.
OptionalmaxConcurrency for target runs (each is an OpenAI-bound chart-review question).
OptionalmetadataExtra metadata stamped on the experiment (git commit, prompt tag, ...) for comparability.
OptionalnumRun every example N times to measure verdict variance — temp-0 is not fully deterministic.
OptionalpatientSugar for splits: ["p<patientId>"] — this cloned patient's examples.
OptionalsettingsBaseline settings overrides — part of the "application version" being evaluated.
OptionalsplitsDataset splits to evaluate, LITERALLY — ["p<patientId>"] runs that patient's examples
regardless of review state (so a freshly-added incoming run can be evaluated before
promotion). Multiple splits are a UNION. Omit (and omit patientId) to run the golden split —
the baseline membership.
LangSmith dataset name, e.g. "Chart Review Run Evals".