Accuracy scores for your category’s fact-checked claims: the numbers behind
the platform’s FactCheck (Accuracy) tab. group_by picks the slice; accuracy
is a ratio 0–1 (accurate / (accurate + inaccurate)). Per-category, so there
is no asset or scope.
group_by: 0–2 of date, model, region, persona, prompt, topic, tag, theme — or one citation (citation can’t be combined with another dimension). Empty → one headline score; ["date"] → the daily accuracy series; two dims (for example, ["model", "date"]) → one row per combination, in a single query.
Row shape follows group_by: a value dimension → { <dim>: { id, name }, … }; ["citation"] → { citation: { url, citation_category }, … } (a citation is a URL, not an id/name); ["date"] → { date, … }.
filter: scopes which responses are counted: a top-level and of single-field leaves over model, topic, region, persona, prompt, tag. One leaf per field (use in for OR within a field); only topic may be negated. This is a narrower grammar than the other v2 reports (no or/not nesting).
New to the v2 reports? See Filtering & concepts for the shared { info, data } shape, grouping, and pagination. FactCheck uses that envelope but has no scope/assets/metrics params and a constrained filter (above). For the inaccurate claims themselves, see FactCheck Claims.
Streaming (SSE) variant (same body, /stream)
POST /v2/reports/factcheck/stream takes the same request body and
returns Server-Sent Events:
one summary event (the info block), then one result event per row.
limit/cursor are ignored; it returns everything by default. Pass
max_results to cap.
Response (text/event-stream)
event: summarydata: { ...the info block... }event: result ← one per row, same shape as data[] abovedata: {"model": {"id": "a1c9…", "name": "ChatGPT"}, "accuracy": 0.986, "accurate": 1980, "inaccurate": 28}