Pass@k Benchmark Estimator: Sampling Grid & Score Curve (2D)

Evaluation harnesses for autonomous agents run each task many times and score it with pass@k — the unbiased probability that at least one of k sampled attempts succeeds. This 2D simulator lays the whole benchmark out as a literal sampling grid — one row per task, one cell per attempt, green for a success and red for a failure — so the raw data behind the score is visible directly, plus a per-task bar showing that row's own pass@k value. Alongside it, a live curve plots the exact closed-form pass@k(k) estimator across every k from 1 to n at once, split into an "all tasks" line and a "clean tasks only" line. Adjust the number of tasks, samples per task, k, the agent's true success rate, and the fraction of tasks contaminated by benchmark leakage to watch the gap between those two curves open up — a direct, hands-on demonstration of why held-out oracles and leakage controls matter more than a single raw benchmark number.