{"schema_version":"papers/paper-detail-v1","title":"Beyond Static Snapshots: A Grounded Evaluation Framework for Language Models at the Agentic Frontier","surface":"papers","opportunity_kernel":{"paper_id":"dce40e63-d79b-4584-82a0-15a20a099508","title":"Beyond Static Snapshots: A Grounded Evaluation Framework for Language Models at the Agentic Frontier","authors":["Jazmia Henry"],"arxiv_id":"2604.17573v1","doi":null,"published_at":"2026-04-19T18:28:32.000Z","score_object":{"overall":{"value":7,"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":"2026-04-21T20:12:14.327Z","fresh_until":"2026-05-21T20:12:14.327Z","is_stale":true,"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":"2026-04-21T20:14:40.082Z","fresh_until":"2026-05-05T20:14:40.082Z","is_stale":true,"source_count":1,"missingness":["reproducibility_results.reproducibility_score","deployability_scores.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":"2026-04-21T20:12:14.327Z","fresh_until":"2026-05-21T20:12:14.327Z","is_stale":true,"source_count":1,"missingness":[]},"market":{"value":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":"2026-04-21T20:32:51.171Z","fresh_until":"2026-05-05T20:32:51.171Z","is_stale":true,"source_count":1,"missingness":[]},"team":{"value":2.7,"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":"2026-04-21T20:12:14.327Z","fresh_until":"2026-05-05T20:12:14.327Z","is_stale":true,"source_count":2,"missingness":[]},"methodology":{"value":3.73,"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":"2026-04-21T20:14:40.082Z","fresh_until":"2026-05-21T20:14:40.082Z","is_stale":true,"source_count":2,"missingness":[]}},"evidence_receipt":{"freshness":"stale","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":"2026-04-21T20:32:51.171Z"},"lineage_hash":"fda639e48c9b09498880166400f1f5090da00730fa1c650233d50923ad952fe7"},"distribution":null,"replication_evidence":[],"author_dna":[]}