OptionaladditionalOptionalauditOptionalconfigTunable levers (models, sampling, loop bounds, toolset, prompts) as a deep-partial override. Missing keys fall back to the accuracy-first package defaults. Production callers resolve this from feature flags at the call site and pass it here.
OptionalcontextPre-hoisted run context (charts + vector stores). When omitted the runner hoists it once for this run. Injecting it avoids a double-fetch on router escalation.
OptionalmodelPer-run model + temperature override. Lets an eval experiment swap the model without editing the committed snapshot or the LangSmith hub. Production leaves it unset -> hub/snapshot model.
OptionalpatientReviewer-selected charts; when set they are the entire candidate set.
OptionalpatientPre-resolved patient context. When omitted the runner resolves it via @repo/db.
Pass it (with context) to reuse a run-scoped resolution across tiers/questions.
OptionalreferenceFrozen "now" for all date reasoning. Production leaves this unset (wall-clock). Eval passes the instant the gold answer was graded against so date-relative questions are reproducible against a frozen data clone instead of drifting to wall-clock.
OptionalreviewRubric that guides grading (from a template's custom_review_criteria_prompt).
OptionaltreatmentOptionaluseUse isolated eval Pinecone namespaces instead of production ones.
This question's own note, stacked on
reviewCriteria. Pass it raw — the runner gates it.