| # PUBLIC prompt-policy oracle-gap evaluation |
|
|
| Training selected **P0** before the held-out bank was opened. All intervals resample the 24 parameter sets and retain their five matched seeds. |
|
|
| | held-out arm | normalized score [95% CI] | deal rate [95% CI] | |
| |---|---:|---:| |
| | canonical P0 | 0.934 [0.903, 0.962] | 0.992 [0.975, 1.000] | |
| | selected P0 | 0.922 [0.883, 0.955] | 0.983 [0.958, 1.000] | |
| | all oracle | 0.791 [0.734, 0.841] | 0.875 [0.817, 0.925] | |
|
|
| ## Oracle-gap closure |
|
|
| - Initial oracle gap (oracle − baseline): -0.143 [-0.197, -0.090] |
| - Prompt-policy gain (selected − baseline): -0.012 [-0.038, 0.013] |
| - Remaining gap (oracle − selected): -0.132 [-0.190, -0.073] |
| - Fraction of initial gap closed: — |
|
|
| Here, `oracle` means the repository's exact sequential omniscient best-response policy. It optimizes each turn from the live state, but it is not a jointly coordinated global optimizer. A negative oracle gap therefore means Opus outperformed this oracle policy, not the global feasible optimum. |
|
|
| Campaign API ledger: $645.20 of the $800 public cap. The Opus private allocation is $1,800; with a $180 discarded-attempt reserve and a $190 robustness allocation, the aggregate program envelope is $2,970. |
|
|
