{"provenance":{"authentication":"derived_from_authenticated_public_release","separately_signed":false,"statement":"This sanitized run projection is derived from the authenticated public release and is not separately signed.","urls":{"manifest":"/data/releases/2026.08-frontier-knowledge-v4/public-manifest.json","manifest_signature":"/data/releases/2026.08-frontier-knowledge-v4/public-manifest-signature.json","public_keys":[{"key_id":"kaelum-bench-public-release-20260718-v1","url":"/data/keys/kaelum-bench-public-release-20260718-v1"}],"run_index":"/data/releases/2026.08-frontier-knowledge-v4/public-run-index.json","self":"/data/runs/6f32c95b-abae-53a0-b9f1-438f8a6ddad0","source_aggregate":"/data/releases/2026.08-frontier-knowledge-v4/fmb-live-runs-capstone.json"}},"run":{"benchmarkId":"FMB-CAP-ACC-03","benchmarkVersion":"0.3.0-dev","campaignId":"74566f95-41ff-521f-b0a5-8fea59044c7b","comparisonGroupId":"accounting-capstone-0.3.0-dev-configuration-specific-v1","comparisonProtocol":"configuration_specific","completedAt":"2026-08-14T07:39:47.037987Z","configuration":{"comparisonProtocol":"configuration_specific","key":"config-b2d14de9f9b48ff5b72675844c905dffc89d278c42a8fbf59672f04bdffeb2f5","materialParameters":[{"name":"allow_fallbacks","value":"false"},{"name":"data_collection","value":"deny"},{"name":"max_attempts_per_task","value":"3"},{"name":"max_concurrency","value":"3"},{"name":"max_generation_metadata_lookups","value":"1"},{"name":"max_output_tokens","value":"44000"},{"name":"max_raw_response_bytes","value":"16777216"},{"name":"max_request_bytes","value":"1048576"},{"name":"metadata_timeout_seconds","value":"60"},{"name":"request_timeout_seconds","value":"300"},{"name":"require_parameters","value":"true"},{"name":"temperature","value":"0"},{"name":"zdr","value":"true"}],"providerRoute":"openrouter/deepinfra/fp4","reasoningEffort":"max","reasoningProfile":"reasoning-max-supported","requestedModelId":"deepseek/deepseek-v4-flash-0731","runtimeMode":"openrouter-tool-free-non-streaming-v8","toolAccess":"none"},"displayVersion":"v0.2","execution":{"attempted":100,"invalidJson":5,"missing":0,"providerFailed":0,"schemaFailed":0,"tasksTotal":100,"valid":95},"globalReleaseId":"c44af57d-9091-45c8-9299-96dee423e175","judge":{"assessmentNotRequiredTasks":0,"assessmentTasks":95,"costUsd":"11.926789875","modelId":"openai/gpt-5.6-sol","providerName":"OpenRouter (Azure EU)","providerRegion":null,"retries":1,"usage":{"inputTokens":277818,"outputTokens":303551,"totalTokens":581369}},"manifest":{"generatedAt":"2026-08-14T12:57:07Z","sha256":"caa004fcccade516afe661a318e69b555844595cbcd726cdc749ff3c64a6378f"},"modelDisplayName":"DeepSeek V4 Flash 0731","providerLabel":"DeepInfra","reasoningEffort":"max","releaseSignature":{"algorithm":"ed25519","keyId":"kaelum-bench-public-release-20260718-v1","messageType":"kaelum.bench.fmb.public_release.v1","payloadSha256":"caa004fcccade516afe661a318e69b555844595cbcd726cdc749ff3c64a6378f","signedFile":"public-manifest.json"},"requestedModelId":"deepseek/deepseek-v4-flash-0731","revisionState":"original","runId":"6f32c95b-abae-53a0-b9f1-438f8a6ddad0","score":{"earned":"787.533333333333","maximum":"1000","percentage":"78.7533"},"scorerId":"fmb-live-scorer-v1","scoringMethod":"hybrid","selectionState":"primary","signedReleaseId":"2459535f-f982-5826-81ed-c8f6d1514773","sourceAggregate":{"campaignId":"74566f95-41ff-521f-b0a5-8fea59044c7b","path":"fmb-live-runs-capstone.json","sha256":"85402a8acfbe9f2ff893e95de5c1e105daada9d13ebf267a87fc67d858020660","sizeBytes":21048},"startedAt":"2026-08-14T06:23:06.760335Z","suiteId":"accounting-capstone","suiteTitle":"Accounting Capstone","testedModelCostUsd":"0.08486304","toolAccess":"none","totalCostUsd":"12.011652915","usage":{"inputTokens":97742,"outputTokens":430296,"totalTokens":528038},"validity":"provisional"},"schema_version":"kaelum-bench-public-run-projection-v1"}