I wrote a method, gave it to Codex, and it passed a Terminal-Bench task that has 59 public runs and zero passes.

59 public runs on this task, across 11 different model and agent configurations. Zero passes. The same model I used, gpt-5.6-sol at max reasoning effort, goes 0 for 5 on it in the public record. One run scored 19 of 19 on the official verifier, inside the 90 minute limit. The task is ico-path-patch on Terminal-Bench 3.0 — binary reverse engineering plus a hot patch, 19 checks, all or nothing. The only thing different about that run is that the agent didn't start on the task. It started by building itself a small service for the task, froze it, then worked the task through that service instead of re-deriving the constraints every few turns. That came from a problem everyone here knows: the longer a run goes, the further the agent's picture of where it is drifts from where it actually is. I wrote a theory about why, from running my own multi-agent system, and turned it into a method after an agent in that system started using the theory to diagnose its own drift unprompted. What I want is for other people to run this task, and there are two ways to do it. Use your own stack. Whatever you've tuned — prompts, orchestration, memory, whatever you've settled on. The task is public, the grader isn't mine, and the result is a single number that doesn't care whose scaffolding produced it. If your setup gets through it with nothing of mine involved, that's a more interesting result than my run, and honestly it would tell me my method isn't the thing doing the work. Or use mine. The steps are written up, it's free, and if it works the score is yours. I'd especially like to see it tried in a domain I've never touched — mine is software-shaped and I have no idea whether this holds up anywhere else. Either way I'd like to hear what happens, including if it does nothing. For what it's worth, it took me four scored attempts before one passed, and there's no run where the agent gets a build phase but no method text — so "any build phase would do" is still a live explanation I can't rule out. Links in the comments.