{"schema_version":"papers/paper-detail-v1","title":"Evaluating the Relevance of Uncertainty Estimators for LLM Hallucination","surface":"papers","opportunity_kernel":{"paper_id":"bafd8975-aa9a-4095-b565-4d0f11af36a4","title":"Evaluating the Relevance of Uncertainty Estimators for LLM Hallucination","authors":["Yedidia Agnimo","Anna Korba","Annabelle Blangero","Nicolas Chesneau","Karteek Alahari"],"arxiv_id":"2605.27016v1","doi":null,"published_at":"2026-05-26T13:34:54.000Z","score_object":{"overall":{"value":4,"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":"2026-05-27T20:11:01.649Z","fresh_until":"2026-06-26T20:11:01.649Z","is_stale":false,"source_count":1,"missingness":[]},"technical":{"value":1.4,"scale":"0-10","confidence":0.55,"confidence_reason":"Backfilled from paper_extraction_scorecards.reconstruction_score.","model_version":"phase0-backfill-v1","pipeline_version":"phase0-kernel-v1","computed_at":"2026-05-27T20:12:21.612Z","fresh_until":"2026-06-10T20:12:21.612Z","is_stale":false,"source_count":1,"missingness":["reproducibility_results.reproducibility_score","deployability_scores.score"]},"commercial":{"value":5,"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":"2026-05-27T20:11:01.649Z","fresh_until":"2026-06-26T20:11:01.649Z","is_stale":false,"source_count":2,"missingness":[]},"market":{"value":5.5,"scale":"0-10","confidence":0.45,"confidence_reason":"Heuristic market score materialized from paper metadata, repo availability, and deployability.","model_version":"phase0-backfill-v1","pipeline_version":"phase0-kernel-v1","computed_at":"2026-05-27T20:33:08.317Z","fresh_until":"2026-06-10T20:33:08.317Z","is_stale":false,"source_count":1,"missingness":[]},"team":{"value":6.4,"scale":"0-10","confidence":0.42,"confidence_reason":"Heuristic fallback from paper author count and extracted affiliations.","model_version":"phase0-backfill-v1","pipeline_version":"phase0-kernel-v1","computed_at":"2026-05-27T20:11:01.649Z","fresh_until":"2026-06-10T20:11:01.649Z","is_stale":false,"source_count":2,"missingness":[]},"methodology":{"value":4.7,"scale":"0-10","confidence":0.82,"confidence_reason":"Backfilled from paper_extraction_scorecards.total_score.","model_version":"phase0-backfill-v1","pipeline_version":"phase0-kernel-v1","computed_at":"2026-05-27T20:12:21.612Z","fresh_until":"2026-06-26T20:12:21.612Z","is_stale":false,"source_count":2,"missingness":[]}},"evidence_receipt":{"freshness":"fresh","proof_status":"unverified","repo_status":"active","references_count":0,"source_count":5,"coverage":0.6667,"missingness":["references","proof_status"],"unresolved_unknowns":["proof verification has not been recorded yet"],"last_verification_at":"2026-05-27T20:33:08.317Z"},"lineage_hash":"2cb5a9599de6e366cd1e1aa93f86ea53c3e7eb1bdae9cb20993de1d008b703ca"},"distribution":null,"replication_evidence":[],"author_dna":[]}