{"$defs":{"canonical_selection":{"additionalProperties":false,"properties":{"contract_version":{"const":"gpqa-results-identity-v1"},"evidence_history":{"items":{"$ref":"#/$defs/evidence_history_item"},"minItems":1,"type":"array"},"precedence":{"items":false,"maxItems":5,"minItems":5,"prefixItems":[{"const":"project_native"},{"const":"direct_strict"},{"const":"direct_compatibility"},{"const":"targeted_recovery"},{"const":"derived_uniform"}],"type":"array"},"score_blind":{"const":true},"selection_timestamp_policy":{"$ref":"#/$defs/nonempty"}},"required":["contract_version","score_blind","selection_timestamp_policy","precedence","evidence_history"],"type":"object"},"configuration_id":{"pattern":"^gpqa_cfg_v1_[0-9a-f]{64}$","type":"string"},"configuration_identity":{"additionalProperties":false,"properties":{"context_policy":{"$ref":"#/$defs/nonempty"},"effort":{"$ref":"#/$defs/nonempty"},"model_snapshot":{"$ref":"#/$defs/nonempty"},"output_policy":{"enum":["source_artifact_policy","source_artifact_policy_plus_uncapped_recovery","no_output_ceiling"]},"prompt_contract":{"enum":["epoch_gpqa_diamond_v1_0_8_task_v0","epoch_gpqa_diamond_source_task_v1","epoch_gpqa_diamond_source_plus_recovery_v1","gpqa_native_answer_choice_v1"]},"protocol":{"$ref":"#/$defs/protocol"},"provider_route":{"$ref":"#/$defs/nonempty"},"scorer_contract":{"const":"inspect_choice_scorer_v1"},"settings":{"oneOf":[{"$ref":"#/$defs/settings_direct"},{"$ref":"#/$defs/settings_derived"},{"$ref":"#/$defs/settings_targeted"},{"$ref":"#/$defs/settings_native"}]}},"required":["context_policy","effort","model_snapshot","output_policy","prompt_contract","protocol","provider_route","scorer_contract","settings"],"type":"object"},"derived_reconciliation":{"additionalProperties":false,"properties":{"paired_original_role":{"$ref":"#/$defs/nonempty"},"published_source_score_percent":{"$ref":"#/$defs/source_percent"},"published_source_score_role":{"$ref":"#/$defs/nonempty"},"source_stderr":{"type":["string","null"]},"status":{"$ref":"#/$defs/nonempty"}},"required":["status","published_source_score_percent","published_source_score_role","source_stderr","paired_original_role"],"type":"object"},"details_derived":{"additionalProperties":false,"properties":{"configuration_date_basis":{"$ref":"#/$defs/nonempty"},"evidence_label":{"$ref":"#/$defs/nonempty"},"reconciliation":{"$ref":"#/$defs/derived_reconciliation"},"recovery":{"$ref":"#/$defs/recovery_uniform"},"score_source":{"$ref":"#/$defs/nonempty"},"source_kind_label":{"$ref":"#/$defs/nonempty"},"source_run_id":{"$ref":"#/$defs/nonempty"}},"required":["configuration_date_basis","evidence_label","source_kind_label","source_run_id","score_source","reconciliation","recovery"],"type":"object"},"details_direct":{"additionalProperties":false,"properties":{"configuration_date_basis":{"$ref":"#/$defs/nonempty"},"evidence_label":{"$ref":"#/$defs/nonempty"},"reconciliation":{"$ref":"#/$defs/direct_reconciliation"},"recovery":{"$ref":"#/$defs/recovery_none"},"score_source":{"$ref":"#/$defs/nonempty"},"source_kind_label":{"$ref":"#/$defs/nonempty"},"source_run_id":{"$ref":"#/$defs/nonempty"}},"required":["configuration_date_basis","evidence_label","source_kind_label","source_run_id","score_source","reconciliation","recovery"],"type":"object"},"details_native":{"additionalProperties":false,"properties":{"configuration_date_basis":{"$ref":"#/$defs/nonempty"},"evidence_label":{"$ref":"#/$defs/nonempty"},"execution":{"$ref":"#/$defs/execution_native"},"reconciliation":{"$ref":"#/$defs/native_reconciliation"},"recovery":{"oneOf":[{"$ref":"#/$defs/recovery_none"},{"$ref":"#/$defs/recovery_native_terminal"},{"$ref":"#/$defs/recovery_native_transport"}]},"score_source":{"$ref":"#/$defs/nonempty"},"source_kind_label":{"$ref":"#/$defs/nonempty"},"source_run_id":{"$ref":"#/$defs/nonempty"}},"required":["configuration_date_basis","evidence_label","source_kind_label","source_run_id","score_source","reconciliation","recovery","execution"],"type":"object"},"details_targeted":{"additionalProperties":false,"properties":{"configuration_date_basis":{"$ref":"#/$defs/nonempty"},"evidence_label":{"$ref":"#/$defs/nonempty"},"reconciliation":{"$ref":"#/$defs/targeted_reconciliation"},"recovery":{"$ref":"#/$defs/recovery_targeted"},"score_source":{"$ref":"#/$defs/nonempty"},"source_kind_label":{"$ref":"#/$defs/nonempty"},"source_run_id":{"$ref":"#/$defs/nonempty"}},"required":["configuration_date_basis","evidence_label","source_kind_label","source_run_id","score_source","reconciliation","recovery"],"type":"object"},"direct_reconciliation":{"additionalProperties":false,"properties":{"difference_percentage_points":{"$ref":"#/$defs/points_six"},"published_source_score_percent":{"$ref":"#/$defs/source_percent"},"published_source_score_role":{"$ref":"#/$defs/nonempty"},"source_stderr":{"type":["string","null"]},"status":{"$ref":"#/$defs/nonempty"}},"required":["status","difference_percentage_points","published_source_score_percent","published_source_score_role","source_stderr"],"type":"object"},"evaluation_run_id":{"pattern":"^gpqa_run_v1_[0-9a-f]{64}$","type":"string"},"evidence_history_item":{"additionalProperties":false,"properties":{"canonical":{"type":"boolean"},"evaluation_run_id":{"$ref":"#/$defs/evaluation_run_id"},"evidence_class":{"enum":["project_native","direct_strict","direct_compatibility","targeted_recovery","derived_uniform"]},"evidence_variant_id":{"$ref":"#/$defs/evidence_variant_id"}},"required":["evaluation_run_id","evidence_variant_id","evidence_class","canonical"],"type":"object"},"evidence_variant_id":{"pattern":"^gpqa_evd_v1_[0-9a-f]{64}$","type":"string"},"execution_native":{"additionalProperties":false,"properties":{"automatic_retries":{"minimum":0,"type":"integer"},"fresh_calls":{"type":"boolean"},"output_ceiling":{"const":"none"}},"required":["automatic_retries","fresh_calls","output_ceiling"],"type":"object"},"native_reconciliation":{"additionalProperties":false,"properties":{"paired_original_role":{"$ref":"#/$defs/nonempty"},"published_source_score_percent":{"const":null},"status":{"const":"not_applicable_independent_run"}},"required":["status","published_source_score_percent","paired_original_role"],"type":"object"},"nonempty":{"minLength":1,"type":"string"},"points_six":{"pattern":"^(?:\\+|-)?(?:0|[1-9][0-9]?|100)\\.[0-9]{6} pp$","type":"string"},"protocol":{"pattern":"^mean@[1-9][0-9]*$","type":"string"},"recovery_native_terminal":{"additionalProperties":false,"properties":{"kind":{"const":"none"},"terminal_no_answer_policy":{"$ref":"#/$defs/nonempty"},"terminal_stop_count":{"minimum":1,"type":"integer"}},"required":["kind","terminal_no_answer_policy","terminal_stop_count"],"type":"object"},"recovery_native_transport":{"additionalProperties":false,"properties":{"kind":{"const":"exact_source_plus_transport_recovery"},"output_ceiling":{"const":"none"},"recovered_source_record_count":{"minimum":1,"type":"integer"},"source_explicit_failure_record_count":{"minimum":0,"type":"integer"},"source_missing_record_count":{"minimum":0,"type":"integer"}},"required":["kind","recovered_source_record_count","source_explicit_failure_record_count","source_missing_record_count","output_ceiling"],"type":"object"},"recovery_none":{"additionalProperties":false,"properties":{"kind":{"const":"none"}},"required":["kind"],"type":"object"},"recovery_targeted":{"additionalProperties":false,"properties":{"configuration_match":{"$ref":"#/$defs/nonempty"},"execution_match":{"$ref":"#/$defs/nonempty"},"kind":{"const":"exact_targeted_transport_recovery"},"model_identity_basis":{"$ref":"#/$defs/nonempty"},"output_ceiling":{"const":"none"},"recovered_pair_count":{"minimum":1,"type":"integer"}},"required":["kind","recovered_pair_count","configuration_match","execution_match","model_identity_basis","output_ceiling"],"type":"object"},"recovery_uniform":{"additionalProperties":false,"properties":{"declared_attempts_per_question":{"minimum":1,"type":"integer"},"kind":{"const":"uniform_epoch_subset"},"omitted_attempt_count":{"minimum":1,"type":"integer"},"selected_attempts_per_question":{"minimum":1,"type":"integer"}},"required":["kind","declared_attempts_per_question","selected_attempts_per_question","omitted_attempt_count"],"type":"object"},"release_id":{"pattern":"^gpqa-clean-results-v[0-9]+\\.[0-9]+\\.[0-9]+$","type":"string"},"row":{"additionalProperties":false,"properties":{"attempts_per_question":{"minimum":1,"type":"integer"},"canonical_evaluation_run_id":{"$ref":"#/$defs/evaluation_run_id"},"canonical_evidence_variant_id":{"$ref":"#/$defs/evidence_variant_id"},"canonical_selection":{"$ref":"#/$defs/canonical_selection"},"configuration_date":{"format":"date","type":"string"},"configuration_date_basis":{"$ref":"#/$defs/nonempty"},"configuration_id":{"$ref":"#/$defs/configuration_id"},"configuration_identity":{"$ref":"#/$defs/configuration_identity"},"details":{"oneOf":[{"$ref":"#/$defs/details_direct"},{"$ref":"#/$defs/details_derived"},{"$ref":"#/$defs/details_targeted"},{"$ref":"#/$defs/details_native"}]},"evidence_class":{"enum":["project_native","direct_strict","direct_compatibility","targeted_recovery","derived_uniform"]},"model_configuration":{"$ref":"#/$defs/nonempty"},"organization":{"$ref":"#/$defs/nonempty"},"protocol":{"$ref":"#/$defs/protocol"},"validation_profile":{"$ref":"#/$defs/nonempty"}},"required":["configuration_id","model_configuration","organization","configuration_identity","configuration_date","configuration_date_basis","canonical_evaluation_run_id","canonical_evidence_variant_id","canonical_selection","protocol","attempts_per_question","evidence_class","validation_profile","details"],"type":"object"},"settings_derived":{"additionalProperties":false,"properties":{"aggregation":{"const":"mean"},"attempts_per_question":{"pattern":"^[1-9][0-9]*$","type":"string"},"declared_attempts_per_question":{"pattern":"^[1-9][0-9]*$","type":"string"},"excluded_epoch_numbers":{"pattern":"^[0-9]+(?:,[0-9]+)*$","type":"string"},"source_model_version":{"$ref":"#/$defs/nonempty"}},"required":["aggregation","attempts_per_question","declared_attempts_per_question","excluded_epoch_numbers","source_model_version"],"type":"object"},"settings_direct":{"additionalProperties":false,"properties":{"aggregation":{"const":"mean"},"attempts_per_question":{"pattern":"^[1-9][0-9]*$","type":"string"},"benchmark_version":{"$ref":"#/$defs/nonempty"},"source_model_version":{"$ref":"#/$defs/nonempty"},"task_version":{"$ref":"#/$defs/nonempty"}},"required":["aggregation","attempts_per_question","benchmark_version","source_model_version","task_version"],"type":"object"},"settings_native":{"additionalProperties":false,"properties":{"aggregation":{"const":"mean"},"attempts_per_question":{"const":"1"},"generation_config_sha256":{"$ref":"#/$defs/sha256"},"profile":{"$ref":"#/$defs/nonempty"},"task_sha256":{"$ref":"#/$defs/sha256"}},"required":["aggregation","attempts_per_question","generation_config_sha256","profile","task_sha256"],"type":"object"},"settings_targeted":{"additionalProperties":false,"properties":{"aggregation":{"const":"mean"},"attempts_per_question":{"pattern":"^[1-9][0-9]*$","type":"string"},"configuration_match":{"$ref":"#/$defs/nonempty"},"execution_match":{"$ref":"#/$defs/nonempty"},"model_identity_basis":{"$ref":"#/$defs/nonempty"},"source_model_version":{"$ref":"#/$defs/nonempty"}},"required":["aggregation","attempts_per_question","configuration_match","execution_match","model_identity_basis","source_model_version"],"type":"object"},"sha256":{"pattern":"^[0-9a-f]{64}$","type":"string"},"source_percent":{"pattern":"^(?:0|[1-9][0-9]?|100)\\.[0-9]{6}%$","type":"string"},"targeted_reconciliation":{"additionalProperties":false,"properties":{"difference_percentage_points":{"$ref":"#/$defs/points_six"},"paired_original_role":{"$ref":"#/$defs/nonempty"},"published_source_score_percent":{"$ref":"#/$defs/source_percent"},"published_source_score_role":{"$ref":"#/$defs/nonempty"},"source_stderr":{"type":["string","null"]},"status":{"$ref":"#/$defs/nonempty"}},"required":["status","difference_percentage_points","published_source_score_percent","published_source_score_role","source_stderr","paired_original_role"],"type":"object"}},"$id":"https://gpqa.ai/schemas/gpqa-clean-results-configurations-v1.json","$schema":"https://json-schema.org/draft/2020-12/schema","additionalProperties":false,"properties":{"configuration_count":{"minimum":1,"type":"integer"},"release_id":{"$ref":"#/$defs/release_id"},"rows":{"items":{"$ref":"#/$defs/row"},"minItems":1,"type":"array"},"schema_version":{"const":"gpqa-clean-results-configurations-v1"}},"required":["schema_version","release_id","configuration_count","rows"],"title":"GPQA Diamond Clean public configuration manifests","type":"object"}
