{"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/02a6a4e8-9ba2-53a7-a5c1-f458ed6e9c64","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":"daf55664-0f64-5efd-82e3-07e31557aa5d","comparisonGroupId":"accounting-capstone-0.3.0-dev-configuration-specific-v1","comparisonProtocol":"configuration_specific","completedAt":"2026-08-14T11:59:01.027623Z","configuration":{"comparisonProtocol":"configuration_specific","key":"config-c77efa5c3bd8fcd155b544e09e1a057bd04d9a5b173c3eba118f5830e4d4ef60","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":"false"}],"providerRoute":"openrouter/alibaba","reasoningEffort":"xhigh","reasoningProfile":"reasoning-max-supported","requestedModelId":"qwen/qwen3.8-max","runtimeMode":"openrouter-tool-free-non-streaming-v8","toolAccess":"none"},"displayVersion":"v0.2","execution":{"attempted":100,"invalidJson":0,"missing":0,"providerFailed":1,"schemaFailed":0,"tasksTotal":100,"valid":99},"globalReleaseId":"c44af57d-9091-45c8-9299-96dee423e175","judge":{"assessmentNotRequiredTasks":0,"assessmentTasks":99,"costUsd":"18.24038205","modelId":"openai/gpt-5.6-sol","providerName":"OpenRouter (Azure EU)","providerRegion":null,"retries":5,"usage":{"inputTokens":291340,"outputTokens":492598,"totalTokens":783938}},"manifest":{"generatedAt":"2026-08-14T12:57:07Z","sha256":"caa004fcccade516afe661a318e69b555844595cbcd726cdc749ff3c64a6378f"},"modelDisplayName":"Qwen3.8 Max","providerLabel":"Alibaba","reasoningEffort":"xhigh","releaseSignature":{"algorithm":"ed25519","keyId":"kaelum-bench-public-release-20260718-v1","messageType":"kaelum.bench.fmb.public_release.v1","payloadSha256":"caa004fcccade516afe661a318e69b555844595cbcd726cdc749ff3c64a6378f","signedFile":"public-manifest.json"},"requestedModelId":"qwen/qwen3.8-max","revisionState":"original","runId":"02a6a4e8-9ba2-53a7-a5c1-f458ed6e9c64","score":{"earned":"357.961904761905","maximum":"1000","percentage":"35.7962"},"scorerId":"fmb-live-scorer-v1","scoringMethod":"hybrid","selectionState":"primary","signedReleaseId":"2459535f-f982-5826-81ed-c8f6d1514773","sourceAggregate":{"campaignId":"daf55664-0f64-5efd-82e3-07e31557aa5d","path":"fmb-live-runs-capstone.json","sha256":"85402a8acfbe9f2ff893e95de5c1e105daada9d13ebf267a87fc67d858020660","sizeBytes":21048},"startedAt":"2026-08-14T10:40:32.404045Z","suiteId":"accounting-capstone","suiteTitle":"Accounting Capstone","testedModelCostUsd":"1.841498","toolAccess":"none","totalCostUsd":"20.08188005","usage":{"inputTokens":86674,"outputTokens":278025,"totalTokens":364699},"validity":"provisional"},"schema_version":"kaelum-bench-public-run-projection-v1"}