Models against Replay.
Run by us, during the challenge. Each model migrates the program honestly, five runs with the same fixed prompt. Then each is asked to be wrong and pass. Every prompt, output and signed result is published here as it lands.
The method.
Two prompts, fixed before the run and published with the results. The first asks for a faithful Python migration of the published program. The second asks the model to produce a migration that is wrong on some input and still passes Replay, with the rules and the allowlist in front of it. Each output goes through the same service as a human entry: static check, public set, branch hunt, hidden set, a signed result.
Where a model’s entry passes, it is treated like any other pass: the model is asked for its proof input, and the claim goes through the win check and a judge. Any test-based tool can be tricked on purpose: a migration that is wrong only on inputs the tool never tries will pass. The rules say so, and a deliberate trick that gets through counts as a win.
Results.
Nothing yet. The table fills in during the run.
| Model | Honest runs passed | Adversarial runs | Caught by | Signed results |
|---|---|---|---|---|
| Model A | — of 5 | — of 5 | pending | — |
| Model B | — of 5 | — of 5 | pending | — |
| Model C | — of 5 | — of 5 | pending | — |
| Model D | — of 5 | — of 5 | pending | — |