KurisutinaRead 2026-09-22. Primary article, Neuroscience & Biobehavioral Reviews 32(2):292–310; PubMed. Public UCL manuscript, retained PDF, text, provenance. Full 31-page manuscript, Methods, Results, Discussion, seven footnotes, mathematical Appendix, six figures and four tables read; graphics and equations checked visually. The publisher's final text and an external supplement were not obtained. The retained manuscript has reporting discrepancies documented below; its acceptance status was not verified.
Factual core
Although indexed as a review, this paper reports a human experiment and computational analyses alongside its review. Nine people with heterogeneous memory disorders and sixteen controls completed 200 probabilistic weather-prediction trials. Every fifty trials they rated each cue's outcome probability and importance to their decisions. Performance improved overall; group differences were not statistically significant. Reported probabilities differentiated cue predictiveness, providing positive evidence of accessible task-relevant knowledge in this sample. Self-rated importance corresponded less clearly to inferred cue use. Among eighteen fitted learning/response variants, BIC favored an associative learner with a constant, individually fitted learning rate and probability matching. Its patient–control rate contrast was marginal, p=.054, one-sided. A dynamic lens analysis separately tracked environmental contingencies and response-derived cue weights. The authors favor a single-system explanation, while acknowledging limited power and alternative reasons for weaker patient performance. These results challenge a blanket claim that probabilistic learning in amnesia occurs without reportable contingency knowledge; they do not establish equal learning between groups or uniquely identify a memory system.
Evidence and inference audit
Data provenance and clinical scope. Methods describes this study's recruitment and experiment, using the procedure of Lagnado et al. 2006. It does not identify these observations as a reanalysis of Meeter or Hopkins data. The nine patients comprise four Korsakoff cases, four post-encephalitis cases, and one right medial-temporal tumor case; medial-temporal involvement was radiologically confirmed in two encephalitis cases and the tumor case. Memory impairment varied from mild to marked. This is not a homogeneous, selectively hippocampal lesion experiment. No original response logs, analysis code, registration, or independent validation sample was obtained. Citations to other experiments are not additional completed readings.
Learning, choosing, and fitting are distinct. The environmental model is an additive logistic function of four binary cue-presence indicators, with no intercept or cue interactions. Associative learning changes these weights through an outcome prediction error. The Bayesian alternative uses a uniform prior on a 7⁴-point weight grid and posterior-mean weights; the grid range is not specified in the manuscript. This particular implementation is not the entire class of Bayesian learning models.
The response transformation raises predicted outcome probabilities to a scaling parameter λ before normalizing. The Appendix correctly gives utilization weights u=λw. At a fixed state, choices alone identify the product, not a unique separation of knowledge strength from response scaling. Constraints on learning trajectories and additional report measures can help distinguish specified models; the algebra does not show that every trajectory is unidentifiable.
Learning-rate and response-scaling parameters are selected by likelihood over the response series, with BIC comparison on those observations. Recursive updates use preceding experience, but fitting their parameters over the series is not a held-out next-choice test. The subsequent lens analysis fixes the rate at .5 and updates a second model from participants' responses. Those utilization weights are descriptive estimates, not independent measurements of internal knowledge or proof of their causal role. Equation 6 scores the expected correctness of an actual categorical choice under the known environment; it is not a proper score for a reported probability forecast.
What explicit knowledge means here. A prompted probability judgment tests reportable contingency knowledge more directly than remembering irrelevant display details. It does not show conscious access at every earlier decision, exact episodic recollection, or that the report causally generated the choice. Importance judgments instead ask about one's own policy and are noisy. A significant late-block contrast alongside nonsignificant earlier contrasts does not alone establish when self-insight emerged. The discussion acknowledges that repeated questioning could influence subsequent learning and cites a separate control experiment; this patient sample contains no unprobed arm.
There is also a conditioning ambiguity. The stated target probabilities .2/.4/.6/.8 average over all patterns containing each card. The model-derived comparison σ(w_j) instead predicts the pattern with only that card present under Equation 1. Our arithmetic check gives .109/.359/.641/.891 from the stated optimal weights −2.10/−.58/.58/2.10. The ratings' wording does not clearly separate these questions. Correct ordering therefore supports task knowledge without establishing calibration to one unambiguous probability target.
Units and strength of evidence. Rating–objective-probability correlations pool 144 patient ratings and 256 control ratings: repeated cards and blocks from nine and sixteen people. Trajectory correlations use df=198 for recursively estimated, serially dependent weights. Their nominal significance should not be read as independent-observation evidence. Participant-level repeated-measures analyses provide additional evidence and must not be discarded simply because these correlations have limits. Report–model correspondence varies substantially between individuals. The paper itself recognizes that null group contrasts may reflect low power or relatively weak control performance. Pooled correlations with memory-test scores do not isolate a within-group causal relation, and fit comparisons cannot exclude correlated contributions from multiple biological systems.
Manuscript-specific reporting checks. Table 4 reverses its Const/Decr prefixes relative to the section headings, rate formulas, and Results. Its decreasing-rate group/matching row prints BIC 6339.72; its stated likelihood, two parameters and N=5000 imply approximately 6217.55. That correction alone does not change the reported overall winner, BIC 5943.44. Table 1's patient delayed-visual mean is printed 85.56, but the nine displayed values average 75.89. Some participant-level rating analyses report df=22 where the full sample would ordinarily give 23, without a stated missing case. These are limits of the retained version, not verified errors in the unavailable publisher version. No empirical results were recomputed from raw data.
Narrow comparison and proposed extraction test
This paper adds a constructive process model and task-relevant report measurements to the Meeter 2006 analysis. Meeter separately predicts a next response from the preceding twenty trials; this paper's fitted learning comparison does not replace that forward test. Neither response profiles nor continuous cue weights uniquely identify a neural memory system. The useful distinction is between an environmentally informed learner, a person's response policy, and what that person can report about either.
Our proposal is to freeze competing learner and policy models using earlier data, then score future choices and separately elicited probabilities under explicit cue-alone and cue-present instructions. Compare the person's report and response histories with matched population baselines, and estimate their incremental value using the same allowed past information. Randomizing the timing of explicit probes would estimate their effect on subsequent learning; a recipient release comparison should condition on the same elicitation history. A functional recipient should be tested both for forecasting the source's decisions and, under a separate objective, using transferred contingencies to act accurately. Success would establish specified functional value, not transfer of an identified memory system.