Pushes one human-graded audit run into the LangSmith eval dataset — one example per question,
keeping the latest response per question (a question re-run within the run dedupes to its newest
grade) and the persisted asked question text (falling back to the template's current text).
New examples land in the incoming split (staged-rollout quarantine) plus the patient and
template axis splits (p<patientId>, t<templateId>) — axis splits keep patient- and
template-scoped experiment selections each a single split name (the splits param is a UNION,
so a combined name would compose worse); exact source-run provenance stays in metadata.
Promotion to golden is a human decision in the LangSmith UI, so pushing can never silently
shift a baseline.
PHI note: question/answer text goes straight to LangSmith (BAA in place) and is never logged.
Pushes one human-graded audit run into the LangSmith eval dataset — one example per question, keeping the latest response per question (a question re-run within the run dedupes to its newest grade) and the persisted asked question text (falling back to the template's current text).
New examples land in the
incomingsplit (staged-rollout quarantine) plus the patient and template axis splits (p<patientId>,t<templateId>) — axis splits keep patient- and template-scoped experiment selections each a single split name (thesplitsparam is a UNION, so a combined name would compose worse); exact source-run provenance stays in metadata. Promotion togoldenis a human decision in the LangSmith UI, so pushing can never silently shift a baseline.PHI note: question/answer text goes straight to LangSmith (BAA in place) and is never logged.