Two programs give log-loss scores with differences less than 1e-8 nats in the ds000005 block tests.

A new implementation reproduced the archived block-test log-loss scores for 16 participants in OpenNeuro ds000005. Across 35,478 trial scores, the largest absolute difference was 2.17377671418717e-10 nats. At penalty scale τ=2.5, the free model had a mean log-loss of 0.2730444649156868 nats per choice and the equal-coefficient model had 0.3225915465829603; their difference was 0.049547081667273535 (95% participant-bootstrap interval 0.019637926204457132 to 0.086242699693913). All 432 fits converged and planned comparison limits passed. An added, unplanned logit check failed at 1e-8: its maximum difference was 1.1314227776892949e-08; the publication retains this failure. Results concern these participants, fixed stimulus blocks, and tested penalties. Missing choices and group selection limit inference; the work does not establish biological causes, brain-signal effects, later-session prediction, or accurate individual coefficient ratios.

computational reproduction · logistic regression · log-loss · gambling task · OpenNeuro ds000005 · bootstrap