{"$defs":{"cohort_summary":{"additionalProperties":false,"properties":{"configuration_count":{"minimum":1,"type":"integer"},"decrease_count":{"minimum":0,"type":"integer"},"definition":{"minLength":1,"type":"string"},"first_quartile_delta":{"$ref":"#/$defs/delta"},"increase_count":{"minimum":0,"type":"integer"},"label":{"minLength":1,"type":"string"},"median_delta":{"$ref":"#/$defs/delta"},"third_quartile_delta":{"$ref":"#/$defs/delta"},"thresholds":{"items":{"$ref":"#/$defs/threshold"},"minItems":1,"type":"array"},"unchanged_count":{"minimum":0,"type":"integer"}},"required":["label","definition","configuration_count","increase_count","decrease_count","unchanged_count","median_delta","first_quartile_delta","third_quartile_delta","thresholds"],"type":"object"},"configuration_id":{"pattern":"^gpqa_cfg_v1_[0-9a-f]{64}$","type":"string"},"delta":{"additionalProperties":false,"properties":{"denominator":{"minimum":1,"type":"integer"},"numerator":{"type":"integer"},"percentage_points":{"$ref":"#/$defs/points"}},"required":["numerator","denominator","percentage_points"],"type":"object"},"percent":{"pattern":"^(?:0|[1-9][0-9]?|100)\\.[0-9]{2}%$","type":"string"},"points":{"pattern":"^(?:\\+|-)?(?:0|[1-9][0-9]?|100)\\.[0-9]{2} pp$","type":"string"},"release_id":{"pattern":"^gpqa-clean-results-v[0-9]+\\.[0-9]+\\.[0-9]+$","type":"string"},"row":{"additionalProperties":false,"properties":{"clean":{"$ref":"#/$defs/score"},"clean_rank":{"minimum":1,"type":"integer"},"configuration_id":{"$ref":"#/$defs/configuration_id"},"delta":{"$ref":"#/$defs/delta"},"model_configuration":{"minLength":1,"type":"string"},"paired_original":{"$ref":"#/$defs/score"},"paired_original_rank":{"minimum":1,"type":"integer"},"protocol":{"pattern":"^mean@[1-9][0-9]*$","type":"string"},"rank_change":{"type":"integer"}},"required":["configuration_id","model_configuration","protocol","paired_original_rank","clean_rank","rank_change","paired_original","clean","delta"],"type":"object"},"score":{"additionalProperties":false,"properties":{"denominator":{"minimum":1,"type":"integer"},"numerator":{"minimum":0,"type":"integer"},"percent":{"$ref":"#/$defs/percent"}},"required":["numerator","denominator","percent"],"type":"object"},"summary_contract":{"additionalProperties":false,"properties":{"quantile_input":{"minLength":1,"type":"string"},"quantile_method":{"const":"exact linear interpolation at (n-1)q"},"threshold_method":{"minLength":1,"type":"string"}},"required":["quantile_method","quantile_input","threshold_method"],"type":"object"},"threshold":{"additionalProperties":false,"properties":{"clean_count":{"minimum":0,"type":"integer"},"crossed_down_count":{"minimum":0,"type":"integer"},"crossed_up_count":{"minimum":0,"type":"integer"},"paired_original_count":{"minimum":0,"type":"integer"},"threshold_percent":{"maximum":100,"minimum":0,"type":"integer"}},"required":["threshold_percent","paired_original_count","clean_count","crossed_up_count","crossed_down_count"],"type":"object"}},"$id":"https://gpqa.ai/schemas/gpqa-clean-results-score-changes-v2.json","$schema":"https://json-schema.org/draft/2020-12/schema","additionalProperties":false,"properties":{"comparison_contract":{"minLength":1,"type":"string"},"release_id":{"$ref":"#/$defs/release_id"},"rows":{"items":{"$ref":"#/$defs/row"},"minItems":1,"type":"array"},"schema_version":{"const":"gpqa-clean-results-score-changes-v2"},"summary":{"additionalProperties":false,"properties":{"all_accepted_configurations":{"$ref":"#/$defs/cohort_summary"},"direct_complete_epoch_artifacts":{"$ref":"#/$defs/cohort_summary"}},"required":["all_accepted_configurations","direct_complete_epoch_artifacts"],"type":"object"},"summary_contract":{"$ref":"#/$defs/summary_contract"}},"required":["schema_version","release_id","comparison_contract","summary_contract","summary","rows"],"title":"GPQA Diamond Clean paired score changes","type":"object"}
