{"$defs":{"artifact":{"additionalProperties":false,"properties":{"bytes":{"minimum":1,"type":"integer"},"path":{"pattern":"^[A-Za-z0-9._/-]+$","type":"string"},"sha256":{"$ref":"#/$defs/sha256"}},"required":["path","bytes","sha256"],"type":"object"},"benchmark":{"additionalProperties":false,"properties":{"excluded_question_count":{"const":9},"exclusions_sha256":{"$ref":"#/$defs/sha256"},"manifest_sha256":{"$ref":"#/$defs/sha256"},"name":{"const":"gpqa-diamond-clean"},"original_question_count":{"const":198},"repository_commit":{"pattern":"^[0-9a-f]{40}$","type":"string"},"retained_question_count":{"const":189},"version":{"type":"string"}},"required":["name","version","repository_commit","manifest_sha256","exclusions_sha256","original_question_count","retained_question_count","excluded_question_count"],"type":"object"},"citation":{"additionalProperties":false,"properties":{"license":{"minLength":1,"type":"string"},"title":{"minLength":1,"type":"string"},"url":{"format":"uri","type":"string"}},"required":["title","url","license"],"type":"object"},"contracts":{"additionalProperties":false,"properties":{"exporter":{"type":"string"},"identity":{"type":"string"},"leaderboard_row":{"type":"string"},"privacy":{"type":"string"},"ranking":{"type":"string"},"release_config":{"type":"string"},"uncertainty":{"type":"string"}},"required":["release_config","exporter","identity","ranking","leaderboard_row","uncertainty","privacy"],"type":"object"},"counts":{"additionalProperties":false,"properties":{"epoch_derived_uniform_configurations":{"minimum":0,"type":"integer"},"epoch_direct_configurations":{"minimum":0,"type":"integer"},"epoch_targeted_recovery_configurations":{"minimum":0,"type":"integer"},"independent_native_configurations":{"minimum":0,"type":"integer"},"ranked_configurations":{"minimum":1,"type":"integer"},"source_artifact_rejections":{"minimum":0,"type":"integer"},"source_ledger_rows":{"minimum":1,"type":"integer"},"superseded_source_rows":{"minimum":0,"type":"integer"},"top_50_coverage_gaps":{"minimum":0,"type":"integer"},"unresolved_diagnostics":{"minimum":0,"type":"integer"}},"required":["source_ledger_rows","ranked_configurations","epoch_direct_configurations","epoch_derived_uniform_configurations","epoch_targeted_recovery_configurations","independent_native_configurations","source_artifact_rejections","superseded_source_rows","top_50_coverage_gaps","unresolved_diagnostics"],"type":"object"},"disclosure":{"additionalProperties":false,"properties":{"minimum_configurations":{"minimum":5,"type":"integer"},"minimum_questions":{"minimum":5,"type":"integer"},"publish_configuration_by_question_matrix":{"const":false},"publish_record_ids":{"const":false},"question_output":{"const":"aggregate_only"}},"required":["question_output","minimum_questions","minimum_configurations","publish_record_ids","publish_configuration_by_question_matrix"],"type":"object"},"ranking":{"additionalProperties":false,"properties":{"aggregation":{"const":"combined_mean_at_n"},"cohort":{"const":"all_fully_accepted_configurations"},"metric":{"const":"clean_accuracy"},"normalize_inference_compute":{"const":false},"tie_policy":{"const":"competition_exact_rational"}},"required":["cohort","aggregation","metric","tie_policy","normalize_inference_compute"],"type":"object"},"release":{"additionalProperties":false,"properties":{"canonical_selection_cutoff_utc":{"format":"date-time","type":"string"},"canonical_selection_timestamp_policy":{"minLength":1,"type":"string"},"data_through":{"format":"date","type":"string"},"id":{"pattern":"^gpqa-clean-results-v[0-9]+\\.[0-9]+\\.[0-9]+$","type":"string"},"public":{"const":true},"publication_status":{"const":"published"},"timestamp_utc":{"format":"date-time","type":"string"},"title":{"minLength":1,"type":"string"},"version":{"pattern":"^[0-9]+\\.[0-9]+\\.[0-9]+$","type":"string"}},"required":["id","version","title","timestamp_utc","canonical_selection_cutoff_utc","canonical_selection_timestamp_policy","data_through","publication_status","public"],"type":"object"},"repository":{"additionalProperties":false,"properties":{"base_commit":{"pattern":"^[0-9a-f]{40}$","type":"string"},"exporter_source_sha256":{"$ref":"#/$defs/sha256"},"note":{"minLength":1,"type":"string"},"uncertainty_builder_source_sha256":{"$ref":"#/$defs/sha256"},"uncertainty_derivation_sha256":{"$ref":"#/$defs/sha256"}},"required":["base_commit","exporter_source_sha256","uncertainty_builder_source_sha256","uncertainty_derivation_sha256","note"],"type":"object"},"sha256":{"pattern":"^[0-9a-f]{64}$","type":"string"},"source_inputs":{"additionalProperties":false,"properties":{"answer_safe_epoch_import_sha256":{"$ref":"#/$defs/sha256"},"epoch_csv_snapshot_sha256":{"$ref":"#/$defs/sha256"},"epoch_source_artifact_bindings_sha256":{"$ref":"#/$defs/sha256"},"independent_native_import_sha256":{"$ref":"#/$defs/sha256"},"question_taxonomy_source_sha256":{"$ref":"#/$defs/sha256"},"targeted_recovery_composite_sha256":{"items":{"$ref":"#/$defs/sha256"},"maxItems":4,"minItems":4,"type":"array","uniqueItems":true},"uniform_recovery_audit_sha256":{"$ref":"#/$defs/sha256"}},"required":["epoch_csv_snapshot_sha256","epoch_source_artifact_bindings_sha256","answer_safe_epoch_import_sha256","uniform_recovery_audit_sha256","targeted_recovery_composite_sha256","independent_native_import_sha256","question_taxonomy_source_sha256"],"type":"object"},"toolchain":{"additionalProperties":false,"properties":{"dependency_lock_sha256":{"$ref":"#/$defs/sha256"},"inspect_ai":{"const":"0.3.240"},"jsonschema":{"const":"4.26.0"},"numpy":{"const":"2.5.0"},"python":{"type":"string"},"pyyaml":{"const":"6.0.3"}},"required":["python","numpy","inspect_ai","pyyaml","jsonschema","dependency_lock_sha256"],"type":"object"},"uncertainty":{"additionalProperties":false,"properties":{"confidence_level":{"const":"95%"},"method":{"const":"synchronized_record_id_cluster_bootstrap"},"replicates":{"minimum":20000,"type":"integer"},"seed":{"minimum":0,"type":"integer"}},"required":["method","seed","replicates","confidence_level"],"type":"object"}},"$id":"https://gpqa.ai/schemas/gpqa-clean-results-release-manifest-v2.json","$schema":"https://json-schema.org/draft/2020-12/schema","additionalProperties":false,"properties":{"artifacts":{"items":{"$ref":"#/$defs/artifact"},"minItems":1,"type":"array"},"benchmark":{"$ref":"#/$defs/benchmark"},"bundle_contract_version":{"const":"gpqa-clean-results-public-bundle-v2"},"citations":{"items":{"$ref":"#/$defs/citation"},"minItems":4,"type":"array"},"contracts":{"$ref":"#/$defs/contracts"},"counts":{"$ref":"#/$defs/counts"},"disclosure":{"$ref":"#/$defs/disclosure"},"independence":{"minLength":1,"type":"string"},"integrity_note":{"minLength":1,"type":"string"},"public_boundary":{"minLength":1,"type":"string"},"ranking":{"$ref":"#/$defs/ranking"},"release":{"$ref":"#/$defs/release"},"repository":{"$ref":"#/$defs/repository"},"schema_version":{"const":"gpqa-clean-results-release-manifest-v2"},"source_inputs":{"$ref":"#/$defs/source_inputs"},"toolchain":{"$ref":"#/$defs/toolchain"},"uncertainty":{"$ref":"#/$defs/uncertainty"}},"required":["schema_version","bundle_contract_version","release","benchmark","contracts","ranking","uncertainty","disclosure","source_inputs","counts","toolchain","repository","citations","artifacts","integrity_note","public_boundary","independence"],"title":"GPQA Diamond Clean results release manifest","type":"object"}
