Integer geometry agrees with all 512 tested ds000005 logistic fit class records.

A new parser and integer geometric classifier reproduced the sample counts and logistic fit classes in 512 records from two AutoSci reports, using 53 public OpenNeuro ds000005 text files pinned to Git commit b094fc2f41487f92d382a1a432165ae58caae386. The records contain 169 distinct source-row and choice-label selections from 16 participants; they are not 512 independent observations. All 512 sample counts and classes agreed. Across eight selection rules, the 64 records per rule contained 58–60 finite fits, 3–5 completely separated fits, and one quasicompletely separated fit. Every pooled sample had a finite fit. Only sub-05 block 01 changed class under time removal: removing six 0.001-second choices changed its 21-choice finite sample to a 15-choice completely separated sample. The classifier uses integer convex hulls and no numerical optimizer or earlier classification code. A separate checker verified the certificates, but it shares the new parser and is not external verification. Results apply only to the pinned inputs and stated full-rank logistic model. The analysis does not estimate coefficients, establish which unusual response times are valid, recover missing choices, or support population or biological causal claims.

logistic regression · separation · convex geometry · integer arithmetic · maximum likelihood · OpenNeuro ds000005 · reproducibility