{"author":"urn:agent:colophon-skillsbench","cohort":{"digest":{"sha256":"f279e3fe7de1308f8177bea06d316b0f6c5d546900f75033396fe55a6cf0d001"},"name":"cohort.json"},"disclosures":{"availability":{"public-exact":492},"closure":{"complete-relative-to-sealed-source":492},"commissioningRequired":false,"conflictsPreserved":0,"evidenceOrigin":{"native-direct":492},"taskRelation":{"exact-digest":492},"timing":{"retrospective-artifacts-only":492}},"limitations":["FLAT POPULATION: 41 statically-admitted tasks, uniform A×5 / B×5 / C×2. No task was selected or dropped on any outcome. The paired estimate is the pre-declared informative subset (14 of 41): C = 0 in every replicate and max(mean A, mean B) > 0. Equivalence margin ±15 pp (150000 ppm).","SCALE: 14 informative tasks against the official Demo-1 floor of 21 units in 13 independence clusters.","HOST-AGENT DEVIATION: the agent ran on the host (Claude Code authenticates itself there), not inside the task image; grading always ran inside the pinned container. Agent-side environment is therefore the host interpreter, not the task's.","SUBJECT MODEL: every cell ran claude-haiku-4-5-20251001. Nothing here generalizes to other models; upstream rescue data shows the skill effect is strongly model-dependent.","SELF-RUN VENUE: the same operator produced and sealed every cell. The checkable artifact chain, not the operator, is what a reader should rely on."],"manifest":{"digest":{"sha256":"822b2f7469dc2e58a3e72eee32688614d296ba20fc381d9a074e3935a68622b3"},"name":"analysis-manifest.json"},"method":{"id":"jinn.benchmarking.method/paired-delta","implementation":{"digest":{"sha256":"d4a071a28bbb1cefa0df37f1c4a7f5c0f510dbcb36bca3b19730dae8a006582f"},"name":"assembly-3.0.json"},"parameters":{"arms":["A-native-skill","B-flat-claude-md"],"control":"C-no-instructions","pairedBy":"task"},"version":"1"},"preregistration":"local-sealed-before-selection","protocol":"https://spec.jinn.network/protocols/benchmarking/v2","results":{"manipulationCheck":{"abMeanPpm":177642,"cCells":82,"cFullPass":7,"cMeanPpm":85366,"upliftPpm":92276},"pairedDelta":{"ci95Ppm":{"lower":-223444,"upper":129159},"meanPpm":-47143,"n":14,"perTask":[{"deltaPpm":-60000,"meanAPpm":53333,"meanBPpm":113333,"replicatesA":5,"replicatesB":5,"samplingVariancePpm":13911,"taskId":"drone-planning-control"},{"deltaPpm":200000,"meanAPpm":200000,"meanBPpm":0,"replicatesA":5,"replicatesB":5,"samplingVariancePpm":40000,"taskId":"edit-pdf"},{"deltaPpm":-200000,"meanAPpm":0,"meanBPpm":200000,"replicatesA":5,"replicatesB":5,"samplingVariancePpm":40000,"taskId":"energy-market-pricing"},{"deltaPpm":200000,"meanAPpm":1000000,"meanBPpm":800000,"replicatesA":5,"replicatesB":5,"samplingVariancePpm":40000,"taskId":"glm-lake-mendota"},{"deltaPpm":0,"meanAPpm":200000,"meanBPpm":200000,"replicatesA":5,"replicatesB":5,"samplingVariancePpm":80000,"taskId":"grid-dispatch-operator"},{"deltaPpm":600000,"meanAPpm":1000000,"meanBPpm":400000,"replicatesA":5,"replicatesB":5,"samplingVariancePpm":60000,"taskId":"hvac-control"},{"deltaPpm":-600000,"meanAPpm":0,"meanBPpm":600000,"replicatesA":5,"replicatesB":5,"samplingVariancePpm":60000,"taskId":"lake-warming-attribution"},{"deltaPpm":-400000,"meanAPpm":200000,"meanBPpm":600000,"replicatesA":5,"replicatesB":5,"samplingVariancePpm":100000,"taskId":"llm-prefix-cache-replay"},{"deltaPpm":-400000,"meanAPpm":0,"meanBPpm":400000,"replicatesA":5,"replicatesB":5,"samplingVariancePpm":60000,"taskId":"mario-coin-counting"},{"deltaPpm":200000,"meanAPpm":600000,"meanBPpm":400000,"replicatesA":5,"replicatesB":5,"samplingVariancePpm":120000,"taskId":"pddl-airport-planning"},{"deltaPpm":0,"meanAPpm":800000,"meanBPpm":800000,"replicatesA":5,"replicatesB":5,"samplingVariancePpm":80000,"taskId":"radar-vital-signs"},{"deltaPpm":0,"meanAPpm":400000,"meanBPpm":400000,"replicatesA":5,"replicatesB":5,"samplingVariancePpm":120000,"taskId":"threejs-structure-parser"},{"deltaPpm":-200000,"meanAPpm":200000,"meanBPpm":400000,"replicatesA":5,"replicatesB":5,"samplingVariancePpm":100000,"taskId":"threejs-to-obj"},{"deltaPpm":0,"meanAPpm":200000,"meanBPpm":200000,"replicatesA":5,"replicatesB":5,"samplingVariancePpm":80000,"taskId":"video-silence-remover"}],"sdPpm":305398,"sePpm":81621,"tCriticalPpm":2160000},"stage":"skillsbench-v1.1-demo1-final","unit":"ppm-of-reward","varianceDecomposition":{"betweenTaskVariancePpm":93268,"heterogeneitySharePpm":238822,"meanSamplingVariancePpm":70994,"taskHeterogeneityPpm":22274}},"subjects":[{"digest":{"sha256":"1bd2dc6cd59df4784e78fca37907000b2f102ec69448a9a233acc9ca7c834470"},"name":"matrix-v2.json"}]}