{"schema_version":"bay-run.canonical-evals.public.v1","availability":false,"artifact_schema_version":null,"required_cases_per_pin":40,"dataset":{"version":null,"digest":null},"evaluated_at":null,"evidence_type":null,"results":[{"pin_id":"route_00857aa05f863c2cdba0e908366b2cca","route_id":"route_00857aa05f863c2cdba0e908366b2cca","model":null,"case_count":null,"score":null},{"pin_id":"route_1c7472e940dc02517f5af93792bf07ee","route_id":"route_1c7472e940dc02517f5af93792bf07ee","model":null,"case_count":null,"score":null},{"pin_id":"route_571826c40685073a99510b1951e60338","route_id":"route_571826c40685073a99510b1951e60338","model":null,"case_count":null,"score":null},{"pin_id":"route_f5411cdb31b03621742a58371fa95732","route_id":"route_f5411cdb31b03621742a58371fa95732","model":null,"case_count":null,"score":null}],"reason":"validated_artifact_absent","sentiment_candidate":{"schema_version":"bay-run.canonical-evals.sentiment-candidate.v1","pin_id":"route_1c7472e940dc02517f5af93792bf07ee","route_id":"route_1c7472e940dc02517f5af93792bf07ee","status":"candidate-unreviewed","reviewed":false,"publication_status":"blocked","case_count":40,"case_file":"evals/candidates/sentiment-tweeteval-test-v1.jsonl","case_file_sha256":"sha256:63c3962c6cc9771ce4d16f66797fa20965a26fe887d8b1d622ca05c26c30dce6","candidate_bundle_file":"evals/candidates/candidate-bundle.v1.json","candidate_bundle_sha256":"sha256:f7502b6f0abd907dc2ec541510eb3a4578a2d7a6e0fa5e6fb139725baea2515f","source":{"dataset":"cardiffnlp/tweet_eval","config":"sentiment","revision":"b3a375baf0f409c77e6bc7aa35102b7b3534f8be","file":"sentiment/test-00000-of-00001.parquet","file_oid":"770b73d16e62f3a93734decccaca43a83c12a1c9","file_sha256":"sha256:f73aea52d43879223019e733b944780f7e9203ee32459ea006f554011379fcdf","split":"test","source_url":"https://huggingface.co/datasets/cardiffnlp/tweet_eval/tree/b3a375baf0f409c77e6bc7aa35102b7b3534f8be","license_note":"Pinned card identifies CC BY 3.0 for sentiment and requires Twitter terms review","license_status":"recorded_not_reviewed","attribution_status":"review_required","platform_terms_status":"review_required","label_mapping":{"0":"NEGATIVE","2":"POSITIVE"},"excluded_source_label":{"1":"NEUTRAL"}},"missing_evidence":[{"code":"rights_attribution_and_platform_terms_review","status":"missing","owner":"human_reviewer","required_action":"Approve the subset's redistribution rights, attribution, and Twitter/Twitter API terms, then record the decision."},{"code":"case_by_case_content_scope_and_pii_review","status":"missing","owner":"human_reviewer","required_action":"Review all 40 cases for PII, duplication, English scope, obvious polarity, and the neutral-label exclusion; approve their use as the held-out public fixture."},{"code":"live_model_outputs_and_metric","status":"missing","owner":"evaluation_operator","required_action":"Run the code-owned sentiment Pin on these exact reviewed 40 cases, retain per-case outputs and receipts, and compute the declared aggregate value from those outputs."},{"code":"evaluation_methodology_and_timestamp","status":"missing","owner":"evaluation_operator","required_action":"Record the deterministic aggregation rule, evaluation method, fixture digest, and a non-future evaluation timestamp."},{"code":"served_model_weight_and_configuration_binding","status":"missing","owner":"evaluation_operator","required_action":"Record the evaluated model, served-weight SHA-256, and configuration hash bound to the run."},{"code":"named_reviewer_and_release_authority_proof","status":"missing","owner":"release_authority","required_action":"Record named reviewer approval and attach the configured bay-run-pin-v1 detached approval envelope over the complete artifact."}],"next_action":"review_then_measure_then_prove"}}