summary-faithfulness

Criteria

keytypeprimitivetier
no_new_factsmust_notllm_judge@12
covers_key_pointsscoredllm_judge@12
no_editorializingmust_notllm_judge@11

summary-faithfulness

**Judges:** whether a summary is faithful to its source — no invented facts/numbers/causality, the key points preserved, no editorial opinion added. Send the summary as `output` and the source as `context` (Contract 05 roles), so the judge is told which text is the authoritative source. **Use when:** you generate summaries whose readers won't check the source (exec digests, meeting notes, news briefs, legal/medical abstracts).

curl -s -X POST $API/v1/rubrics -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d @templates/summary-faithfulness/rubric.json
# typed roles: the summary is the output; the source travels as context
curl -s -X POST $API/v1/verify -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d '{"rubric_id":"<id>","submission":{"output":"<the summary>","context":[{"label":"source","value":"<the source document>"}]},"options":{"wait_ms":45000}}'

MCP: `verify_submit` with the summary as `output` and the source document as `context`. The legacy `submission.inline` shape still works unchanged.