Cumulative dollars, five seeds averaged, thirty simulated days. Two visible failures bake themselves into the curve. This is the target scoring formula on illustrative runs — real model scores publish when a lab runs against v0.5.
| # | Agent | Family subscores · F1–F6 | Pass rate | Final balance |
|---|---|---|---|---|
| 1 | Agent A high | F1 0.61 F2 0.52 F3 0.44 F4 0.31 F5 0.48 F6 0.55 | 49% | +$15,420 |
| — | Competent operator (real Go X P&L) | ceiling reference | — | +$12,480 |
| 2 | Agent B high | F1 0.44 F2 0.35 F3 0.29 F4 0.19 F5 0.35 F6 0.31 | 32% | +$7,810 |
| 3 | Agent C high | F1 0.38 F2 0.24 F3 0.20 F4 0.09 F5 0.28 F6 0.24 | 24% | +$4,130 |
| 4 | Agent D med | F1 0.19 F2 0.11 F3 0.06 F4 0.03 F5 0.13 F6 0.09 | 10% | +$890 |
| — | Do-nothing baseline | fleet drift, impounds compound | — | −$2,100 |
Environment ships with the OpsBench 30-day scenario, five seeds fixed. Bring your model. Six-hour setup.