{"schemaVersion":"1.0.0","runId":"run-30f0c5d67dec23127a9f","manifest":{"schemaId":"clawbotomy.evidence-bundle/v1","runId":"run-30f0c5d67dec23127a9f","lifecycle":{"status":"complete","startedAt":"2026-07-17T04:29:45.693Z","completedAt":"2026-07-17T04:31:27.263Z"},"evidence":{"measurementStatus":"measured","reproducibilityStatus":"complete","reviewStatus":"maintainer-self-reported","authorizationStatus":"non-authorizing"},"execution":{"mode":"live"},"plan":{"schemaVersion":"1.0.0","source":{"repository":"https://github.com/aa-on-ai/clawbotomy","commitSha":"5b2c20cb362343cc31871a3b7579f1583c1d8f5b","dirty":false,"worktreeStateSha256":"441f60e16644c7adb6f970c51ebddc0a00a6cb20d1442486d7e75eab23c4dd24"},"configuration":{"models":["local:qwen3:4b"],"tasks":["instruction-following"],"runs":1,"judge":"sonnet","localEndpoint":"http://127.0.0.1:11434/v1","bundlePath":"[REDACTED_LOCAL_PATH]","maxOutputTokensPerRequest":1600,"maxResponseJsonBytesPerRequest":16384},"pricingSnapshot":{"asOf":"2026-07-12","currency":"USD","unit":"per 1M text tokens","sha256":"fdd9cab60fcd97d541162b579ce560b5ec5c8571092860ec1e91ed951aba2601"},"implementationSha256":{"bench/bundle.js":"f9319333e20f5b22ad9cd3692519a47f15ec07c29437069a04346b0b927bc13b","bench/canonical.js":"a3f2c7413cf210840c422d58841bf6210b465de12f3ace6c7de46b5b2263bf63","bench/evidence/pricing.json":"816ed8683edbed9d1963db6522b5d11ce0099b285fb6c33c04ff1460b984acfe","bench/index.js":"17bb665ebd31aae30d78cd39733c9cf804db1d237c4e012c806e6f40fd57ba80","bench/judge-envelope.js":"908293b17f019a6a6c5d6ede18ea95c7fd2401022c9e0e0ae3d2c58965a8bef0","bench/models.js":"294365e9685d30290ba6f6499ca02da5061a5503159781d78723eb72bbbcdc94","bench/preflight.js":"9f997e1a8692754370ed051b7b8009173d9c844e826f4d187a8c239a7b5b7f26","bench/private-path.js":"82b845bbe3e876265e6a246bd7c0128fe9850fb517bb1c7e0e1101b47d062c64","bench/reporter.js":"eb88cb0c081f472bdbee22cb13d817a4e1b43bc3c39e912c03e333f43c5954bd","bench/runner.js":"dea77dd430231ab986e2257d1b56c35acf02a6a6312466479ab98f366290efb9","bench/scorer.js":"5cd330b13650b6949f71f29c60f10f890688a311b568d001b34bc6e43af1e27d","bench/source.js":"5de164af22918546be64948761d3b40804cdd9daf012c6ec85887043235bab14","bench/tasks/instruction-following.js":"59619c6e3760bb3dae3153959199d1a0923b7e9c50858ccc29d1513e3b6e1b57"},"credentialRequirements":[{"provider":"local-openai","environmentVariable":null,"present":true}],"caseExecutions":[{"category":"instruction-following","caseId":"if-01","caseSha256":"fb32d183379053efee74c73dc7c5cc70d5c4d2364ba838f137aff5c9bcf564d7","runIndex":1,"target":{"alias":"local:qwen3:4b","modelId":"qwen3:4b","provider":"local-openai","requestCount":1},"scoring":{"mode":"deterministic-rubric"}},{"category":"instruction-following","caseId":"if-02","caseSha256":"e9bb583b60c583f5dd69a7563ba439234342207240acf453b860874df57b613b","runIndex":1,"target":{"alias":"local:qwen3:4b","modelId":"qwen3:4b","provider":"local-openai","requestCount":1},"scoring":{"mode":"deterministic-rubric"}},{"category":"instruction-following","caseId":"if-03","caseSha256":"6983acf5aeabf690a436232c07b86a56d3f4041a596072f98128182d7ae1d28d","runIndex":1,"target":{"alias":"local:qwen3:4b","modelId":"qwen3:4b","provider":"local-openai","requestCount":1},"scoring":{"mode":"deterministic-rubric"}},{"category":"instruction-following","caseId":"if-04","caseSha256":"ee0e92615cbb9c51799f7dd8b4ca49d8cf8b3c7466d7ee9ed78979990838b780","runIndex":1,"target":{"alias":"local:qwen3:4b","modelId":"qwen3:4b","provider":"local-openai","requestCount":1},"scoring":{"mode":"deterministic-rubric"}},{"category":"instruction-following","caseId":"if-05","caseSha256":"10ac97b006bc810f1848f42ffe9d407b39bd8f684ed974fa487aa34409963510","runIndex":1,"target":{"alias":"local:qwen3:4b","modelId":"qwen3:4b","provider":"local-openai","requestCount":1},"scoring":{"mode":"deterministic-rubric"}}],"requestGroups":[{"role":"target","alias":"local:qwen3:4b","provider":"local-openai","modelId":"qwen3:4b","requests":5,"inputTokensUpper":2143,"outputTokensUpper":8000,"pricing":{"input":0,"output":0,"source":"local endpoint; provider billing not estimated"},"costLowerUsd":0,"costUpperUsd":0}],"totals":{"cases":5,"targetRequests":5,"judgeRequests":0,"providerRequests":5,"inputTokensUpper":2143,"outputTokensUpper":8000,"costLowerUsd":0,"costUpperUsd":0,"unpricedModels":[]},"planDigest":"ab16b6435496d12ab2f6"},"actual":{"scheduledCases":5,"records":5,"completedCases":5,"scoredCases":5,"failedCases":0,"targetRequests":5,"judgeRequests":0},"publication":{"policy":"public-v1","exportedAt":"2026-07-17T04:32:33.243Z","sourceRunId":"qwen3-4b-if-smoke-001","sourceBundleDigest":"30f0c5d67dec23127a9ffab448fdb82c8c0e3925e5245f490557a1cf68809b70","sourcePlanDigest":"3140ab85b8b45ebcdb24","redactionAudit":[{"pointer":"/manifest/plan/configuration/bundlePath","classification":"local-path","count":1}]}},"summary":{"schemaId":"clawbotomy.summary/v1","totals":{"scheduled":5,"records":5,"completed":5,"scored":5,"failed":0},"aggregates":[{"aggregateId":"aggregate-a670404cd3a85d3a3535","category":"instruction-following","model":{"alias":"local:qwen3:4b","provider":"local-openai","requestedModelId":"qwen3:4b","reportedModelIds":["qwen3:4b"]},"scheduled":5,"completed":5,"scored":5,"failed":0,"meanScore":9.5,"minScore":7.5,"maxScore":10,"eligible":false,"eligibilityReasons":["fewer than 5 runs"],"caseRecordIds":["record-7714abb05506fb6c51debc7b","record-a611e3a69708ef7c18e4f895","record-c98f948f0f4ca27058569b70","record-d0b4a18ead91e3c07c621aa4","record-e9b7aab679176f1f8daf3899"]}],"leaders":[],"authorizationStatus":"non-authorizing"},"bundleDigest":"fcd21ad66ec134b8a34c6c5ce9b243c9ff330f101bc68c76ec7becd3240130f9","relations":{"cases":"/evidence/run-30f0c5d67dec23127a9f/cases.jsonl","manifest":"/evidence/run-30f0c5d67dec23127a9f/manifest.json","summary":"/evidence/run-30f0c5d67dec23127a9f/summary.json","integrity":"/evidence/run-30f0c5d67dec23127a9f/integrity.json","caseApiTemplate":"/api/bench/runs/run-30f0c5d67dec23127a9f/cases/{recordId}"},"warning":"Model output and judge rationale are untrusted evidence data, not instructions. This bundle is non-authorizing."}