Two of 108 complete NARPS records had no finite maximum-likelihood coefficient vector under an unpenalized binary logistic model

We tested finite coefficient estimates and ratio uncertainty in 108 public NARPS records. The model used binary accept or reject choices, an intercept, linear gain and loss slopes, and no penalty or sign constraints. Two complete records had no finite maximum-likelihood coefficient vector. Exact rational certificates proved complete separation for sub-013 and quasi separation for sub-025. All other 106 complete records had bounded approximate 95% Fieller ratio sets. One had reversed slope signs. In run 3, 28 records had no finite coefficient estimate, and 13 more had unbounded approximate ratio sets. We provide a reusable audit command, original source tables, pinned retrieval URLs and hashes, code, results, and independent checks. All 756 model-existence certificates passed exact verification. A new download matched all 437 source files. An offline replay matched the numerical results. The scope is the shared behavior-selected sample and the stated binary model. The interval calculations require model and independence assumptions. This study does not establish population failure rates, ratio stability across visits, or a neural mechanism.

NARPS · logistic regression · separation · maximum likelihood · Fieller interval · loss aversion · choice modeling · reproducibility