{"schema_version":"papers/paper-detail-v1","title":"CalArena: A Large-Scale Post-Hoc Calibration Benchmark","surface":"papers","opportunity_kernel":{"paper_id":"506d9a2c-7742-418e-8aba-84fba3cea8fa","title":"CalArena: A Large-Scale Post-Hoc Calibration Benchmark","authors":["Eugène Berta","David Holzmüller","Francis Bach","Michael I. Jordan"],"arxiv_id":"2605.30188v1","doi":null,"published_at":"2026-05-28T16:31:36.000Z","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":"2026-05-29T20:12:08.123Z","fresh_until":"2026-06-28T20:12:08.123Z","is_stale":false,"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":"2026-05-29T20:34:23.386Z","fresh_until":"2026-06-12T20:34:23.386Z","is_stale":false,"source_count":2,"missingness":["reproducibility_results.reproducibility_score"]},"commercial":{"value":4,"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":"2026-05-29T20:12:08.123Z","fresh_until":"2026-06-28T20:12:08.123Z","is_stale":false,"source_count":1,"missingness":[]},"market":{"value":2.2,"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":"2026-05-29T20:36:57.989Z","fresh_until":"2026-06-12T20:36:57.989Z","is_stale":false,"source_count":1,"missingness":[]},"team":{"value":3.6,"scale":"0-10","confidence":0.28,"confidence_reason":"Heuristic fallback from paper author count.","model_version":"phase0-backfill-v1","pipeline_version":"phase0-kernel-v1","computed_at":"2026-05-29T20:12:08.123Z","fresh_until":"2026-06-12T20:12:08.123Z","is_stale":false,"source_count":1,"missingness":["paper_author_affiliations"]},"methodology":{"value":3.9,"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":"2026-05-29T20:18:52.227Z","fresh_until":"2026-06-28T20:18:52.227Z","is_stale":false,"source_count":2,"missingness":[]}},"evidence_receipt":{"freshness":"fresh","proof_status":"unverified","repo_status":"missing","references_count":0,"source_count":3,"coverage":0.5,"missingness":["repo_url","references","proof_status"],"unresolved_unknowns":["proof verification has not been recorded yet"],"last_verification_at":"2026-05-29T20:36:57.989Z"},"lineage_hash":"1580c2c1346ea999d737bcd0cf3984a9da26c7a15355e210879f6afd64dbec40"},"distribution":null,"replication_evidence":[],"author_dna":[]}