# Per-case accepted-result comparison **Tested October 30, 2026. Observational CLI-adapter V2. All twelve first attempts retained; no repairs.** | Run | Task | Repeat | Lane | Accepted | Scores A/B | Seconds | Reported input/output tokens | Task calls | Handoffs | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | R01 | wallet | 2 | team | True | 100/100 | 72.548 | 147,919 | 18 | 4 | | R02 | wallet | 2 | solo | True | 100/100 | 65.801 | 68,510 | 18 | 0 | | R03 | wallet | 2 | hybrid | True | 100/100 | 71.082 | 65,032 | 17 | 0 | | R04 | subscription | 1 | solo | True | 90/90 | 63.972 | 59,772 | 13 | 0 | | R05 | subscription | 1 | hybrid | True | 95/95 | 56.807 | 63,648 | 14 | 0 | | R06 | subscription | 1 | team | True | 95/95 | 86.588 | 204,410 | 19 | 4 | | R07 | wallet | 1 | hybrid | True | 100/100 | 60.348 | 68,488 | 18 | 0 | | R08 | wallet | 1 | solo | True | 100/100 | 67.698 | 64,703 | 18 | 0 | | R09 | wallet | 1 | team | True | 100/100 | 110.478 | 167,932 | 20 | 4 | | R10 | subscription | 2 | hybrid | True | 95/95 | 58.244 | 63,474 | 14 | 0 | | R11 | subscription | 2 | team | True | 95/95 | 82.517 | 152,543 | 14 | 4 | | R12 | subscription | 2 | solo | True | 95/95 | 57.543 | 58,749 | 10 | 0 | ## Per-task summaries | Task | Lane | Accepted | Mean score | Median seconds | Median reported tokens | | --- | --- | --- | --- | --- | --- | | subscription | solo | 2/2 | 92.5 | 60.757 | 59,260.5 | | subscription | team | 2/2 | 95 | 84.553 | 178,476.5 | | subscription | hybrid | 2/2 | 95 | 57.525 | 63,561.0 | | wallet | solo | 2/2 | 100 | 66.750 | 66,606.5 | | wallet | team | 2/2 | 100 | 91.513 | 157,925.5 | | wallet | hybrid | 2/2 | 100 | 65.715 | 66,760.0 | ## Observable work per accepted result All attempts count in each numerator. Dollars remain unknown. | Lane | Total reported tokens / accepted | Total seconds / accepted | Economic cost / accepted | | --- | --- | --- | --- | | solo | 62,933.5 | 63.754 | Unknown | | team | 168,201.0 | 88.033 | Unknown | | hybrid | 65,160.5 | 61.620 | Unknown | ## Failure and qualification notes No critical factual/arithmetic failure was found. Both graders deducted source-date points from all six subscription answers; one solo answer also omitted the unknown-publication-date statement. These are minor reporting omissions, not invented expensive mistakes. Wallet answers all scored 100. Human diagnosis and cleanup time were not measured. All four hybrid runs chose zero specialists. All four team runs had two specialists with overlapping sessions, yet took longer than the paired solo run. No task-call, artifact-length, session-integrity, model-tag or forbidden-action violation was recorded. Reported tokens sum captured response events across every participant process, including cached input and startup events. They are not billing receipts or subscription burn. Resumed coordinator counters were reconciled rather than added twice. The original controlled V1 ledger remains not run. Two task clusters, two repetitions, procedural isolation, same-model automated grading and this particular adapter constrain the findings. No publication, paid API-key request or purchase occurred.