A test martingale at threshold 20 bounds false rejection by 5% under a known conditional stop law and predictable response scores

We give a Python tool to test current response information about future stops. Researchers must set each score before they know its next label. The conditional stop probability must be correct under the null. A proof gives a 5% false rejection bound, including repeated inspection of one fixed test. The general probability theorem is established. Public AOMIC PIOP2 event files set the benchmark size at 100 trials and 226 participants. We used 5,000 simulated studies per condition. In five settings with correct laws, detection rates were 2.18% to 2.86% for responses with only past inputs. An ordinary future stop contrast gave 83.22% detections under independent stops with a response from the past stop count. The tests have different nulls. An incorrect adaptive probability gave 100% detections without future response input. The new work is the checked tool and its benchmark. Power was low in some conditions with one participant. The package keeps code, proof, 233 pinned source files, simulation records, and reproduction instructions. An independent rerun gave identical numerical arrays and summaries. We make no calibrated inference about recorded human responses because their required assignment law is unknown.

test martingale · sequential testing · conditional assignment · stop signal task · simulation · false rejection · predictable score