Turn a bug report into a replay.
Nemotron Super reads a player's bug report and writes an oracle: exact conditions over engine events that the crowd can test every round against. Rounds that trip the oracle become replayable reproductions.
Fresh-play experiment
Rare report
“The player floats upward while blocking after being released by the rival near the wall, then flies out of bounds.”
Chooser model: gemma4:e4b-mlx
| Arm | Reproduced | Calls | Fallbacks |
|---|---|---|---|
| recorded build | 0 of 20 | 383 | 89 (23%) |
| shipped build | 0 of 20 | 159 | 53 (33%) |
| scripted control | 0 of 20 | 0 | 0 |
Fallback rates: recorded build 89 of 383 calls (23%), shipped build 53 of 159 calls (33%).
Tested with a local small model (gemma4:e4b-mlx), not Nemotron, across 20 fresh seeds for one bug report; this is one report, one run, not a success rate. The shipped-build arm tests the current game where the bug is fixed, so a fixed bug should reproduce 0 times. Base rate: that bug is rare (4 of 2,831 rounds; about 2.8% chance of one hit in 20 rounds, so 0 of 20 shows nothing).
High fallback rate: failed model responses led to wait actions, so this run may not represent a functioning player.
Common report
“The player is knocked down at least once in the first 30 seconds.”
Chooser model: gemma4:e4b-mlx
| Arm | Reproduced | Calls | Fallbacks |
|---|---|---|---|
| recorded build | 19 of 20 | 162 | 42 (26%) |
| shipped build | 19 of 20 | 116 | 31 (27%) |
| scripted control | 20 of 20 | 0 | 0 |
Fallback rates: recorded build 42 of 162 calls (26%), shipped build 31 of 116 calls (27%).
Tested with a local small model (gemma4:e4b-mlx), not Nemotron, across 20 fresh seeds; this is one report, one run, not a success rate. What the arms mean: the shipped and recorded builds should reproduce at similar rates because it is not a bug. The scripted control arm provides a baseline without model calls.
High fallback rate: failed model responses led to wait actions, so this run may not represent a functioning player.