{"schema_version":"papers/paper-detail-v1","title":"Automated Rubrics for Reliable Evaluation of Medical Dialogue Systems","surface":"papers","opportunity_kernel":{"paper_id":"ea4ac747-d13e-46e4-9a8f-c9c2e20df93e","title":"Automated Rubrics for Reliable Evaluation of Medical Dialogue Systems","authors":[],"arxiv_id":"2601.15161v1","doi":null,"published_at":"2026-01-21T16:40:41.000Z","score_object":{"overall":{"value":8,"scale":"0-10","confidence":0.85,"confidence_reason":"Backfilled from persisted papers.viability_score.","model_version":"phase0-backfill-v1","pipeline_version":"phase0-kernel-v1","computed_at":"2026-04-02T02:30:40.136Z","fresh_until":"2026-05-02T02:30:40.136Z","is_stale":true,"source_count":1,"missingness":[]},"technical":{"value":2.4,"scale":"0-10","confidence":0.7,"confidence_reason":"Backfilled from deployability_scores.score.","model_version":"phase0-backfill-v1","pipeline_version":"phase0-kernel-v1","computed_at":"2026-03-17T21:43:58.792Z","fresh_until":"2026-03-31T21:43:58.792Z","is_stale":true,"source_count":1,"missingness":[]},"commercial":{"value":0,"scale":"0-10","confidence":0.75,"confidence_reason":"Backfilled from persisted commercial_flags and repo availability.","model_version":"phase0-backfill-v1","pipeline_version":"phase0-kernel-v1","computed_at":"2026-04-02T02:30:40.136Z","fresh_until":"2026-05-02T02:30:40.136Z","is_stale":true,"source_count":1,"missingness":[]},"market":{"value":0,"scale":"0-10","confidence":0.15,"confidence_reason":"No persisted distribution_readiness_scores row was available; marked low confidence.","model_version":"phase0-backfill-v1","pipeline_version":"phase0-kernel-v1","computed_at":"2026-04-02T02:30:40.136Z","fresh_until":"2026-04-16T02:30:40.136Z","is_stale":true,"source_count":0,"missingness":["distribution_readiness_scores.score"]},"team":{"value":10,"scale":"0-10","confidence":0.15,"confidence_reason":"No persisted team-quality evidence was available; marked low confidence.","model_version":"phase0-backfill-v1","pipeline_version":"phase0-kernel-v1","computed_at":"2026-04-02T02:30:40.136Z","fresh_until":"2026-04-16T02:30:40.136Z","is_stale":true,"source_count":0,"missingness":["engineer_profiles.builder_score","author_startups"]},"methodology":{"value":0,"scale":"0-10","confidence":0.15,"confidence_reason":"No persisted methodology score source was available; marked low confidence.","model_version":"phase0-backfill-v1","pipeline_version":"phase0-kernel-v1","computed_at":"2026-04-02T02:30:40.136Z","fresh_until":"2026-05-02T02:30:40.136Z","is_stale":true,"source_count":0,"missingness":["paper_extraction_scorecards.total_score","paper_extraction_scorecards.standard_extraction_score"]}},"evidence_receipt":{"freshness":"stale","proof_status":"unverified","repo_status":"missing","references_count":0,"source_count":0,"coverage":0.3333,"missingness":["repo_url","references","distribution_readiness_scores","paper_extraction_scorecards"],"unresolved_unknowns":["distribution readiness has not been computed yet"],"last_verification_at":"2026-03-17T21:43:58.792Z"},"lineage_hash":"1b506e25884f90f21a806d5c2191e2e2cf494b666b125e73e9612251f400976c"},"distribution":null,"replication_evidence":[],"author_dna":[]}