KurisutinaRead 2026-09-22; saved at the user's request to finish the session.
Citation. Michael J. Wolff, Janina Jochim, Elkan G. Akyürek and Mark G. Stokes. Dynamic hidden states underlying working-memory-guided behavior. Nature Neuroscience 20, 864–871. DOI, PMC5446784. Published 17 April 2017. Primary human EEG/behavioral experiments.
Reading scope. Read the complete repository main text, Results, Discussion, Online Methods and availability statements. Checked all six main figures visually in the 41-page institutional manuscript, plus the equations and consequential method details on pages30–33. Read all four publisher supplementary figures/captions and the substantive ten-page reporting checklist. Read the complete publisher MATLAB decoding helper and its later GitHub counterpart; inspected the first130lines of the relevant figure script. No author code was executed. The institutional PDF contains 36 main/reference/method pages, a divider and four supplemental pages; it was not completely line-compared with the repository XML or final publisher main. Bibliographic entries were not independently read as cited studies. The readable XML extraction repeats captions and contains bibliography labels rather than complete reference entries; the original XML and PDF retain those entries.
Local manuscript, PDF text, repository XML, readable main, publisher supplement, reporting checklist, publisher code and provenance manifest. Public accessibility does not establish an unrestricted redistribution license.
Factual core, under200words. Participants remember two grating orientations. Experiment1 uses a valid retrospective cue to select the tested item; a fixed visual impulse precedes the test. Its evoked EEG contains orientation-related information for the cued item, exceeding the uncued item's measure. Experiment2 tests both items sequentially: the first impulse carries information about both the immediately relevant and later-tested item, while ordinary delay decoding favors the prioritized item. Trial-wise impulse scores relate to subsequent accuracy; selected psychometric comparisons also relate to slope. A separate behavioral experiment finds no detected performance effect of varying impulse timing. These results support task-dependent, orientation-specific structure in responses to a common input, beyond broad category decoding. They do not directly measure synapses, establish that discarded items have been erased, or provide a calibrated estimate of an unknown orientation. The public analysis aligns each trial to its known orientation when constructing distance-tuning templates. Relevant data and code are advertised publicly; neither an empirical reproduction nor a blind extraction test was conducted in this reading.
Design and positive controls. Experiment1 retains30 of34 participants; Experiment2 retains19 of20; Experiment3 retains20 of26. Exclusion rules include performance below60% and/or excessive eye movements, with criteria described as established in advance in the checklist. No power calculation or investigator blinding is reported. Experiment1 has1,344 trials over approximately3hours; Experiment2 has864 trials per session on two days1–2weeks apart, analyzed within person/session, not as38 independent people. Experiment3 has280 trials over approximately30minutes.
The cue indicates which side matters without providing the angle. Thus a cue-only recipient does not automatically know the target orientation. Within Experiment1 the two items have comparable presentation histories and a direct cued–uncued neural comparison is reported: average100–500ms difference p=.028, with a corrected temporal cluster as additional evidence. This is stronger than comparing one significant result with one nonsignificant result. Experiment2's retained but later-tested orientation provides a useful contrast with the discarded item of Experiment1, although the experiments also differ in timing, impulse appearance and task structure.
What the decoder measures. The helper computes leave-one-trial-out Mahalanobis distances between the test pattern and training patterns averaged in overlapping orientation bins. The bins are centered relative to the known test orientation. A cosine-weighted mean summarizes whether nearby orientations have more similar activity patterns. Training means and covariance exclude the test trial; using the known label to evaluate tuning is not, by itself, evidence of invalid neuroscience. It does mean the reported “decoding accuracy” is a continuous label-aligned statistic, not a proportion of correctly recovered angles or a probability distribution supplied to a recipient. A blind decoder would require templates or a model applicable without the unknown test label and a separate held-out error/calibration evaluation.
Preprocessing includes filtering, baseline correction and within-pattern channel demeaning. The temporal curves are smoothed with a Gaussian of SD16ms. Leave-one-trial-out fitting uses other trials throughout the session, including later ones; it does not establish prospective online decoding. Filtering details and smoothing must be accounted for before assigning a real-time availability claim. Visual EOG/artifact rejection is reported, but no eye-signal-only orientation-decoding control is established by this reading.
Mechanistic limits. Nondetection of discarded content does not prove the corresponding information has been eliminated from every retained state. Discarded items are not later tested to establish absence of recoverability. Failed encoding-to-impulse cross-generalization constrains the tested readouts, not every common representational mechanism. Likewise, the absence of a correlation with alpha lateralization does not prove attention is irrelevant. The positive impulse result is compatible with a changed input–output response profile; direct synaptic storage remains unmeasured.
Experiment3's null is useful evidence about the tested behavioral contrast, not an equivalence test demonstrating a perfectly passive measurement. It changes impulse duration and uses a different probe appearance from Experiment1, has only20 retained participants, and measures final discrimination accuracy. Other source effects or small effects remain possible.
Behavior and statistics. Trial-wise scores are median-split, followed by within-person high–low accuracy comparisons. These are associations using a known orientation, not held-out forecasts of a new person's behavior. The psychometric slope combines uncertainty from memory, perception and decision; it is not an isolated trace-strength measurement. In the displayed model y=λ+(1−2λ)Φ(β(x−α)), λ is the lower asymptote. Under a uniform random-guess mixture interpretation, the mixture's guess probability is2λ, not λ. One-tailed slope/asymptote tests follow observed accuracy effects; multiple time-window hypotheses were not jointly corrected. Temporal clusters are corrected, and group tests use participants as the unit. Sign-flip testing still requires appropriate exchangeability/symmetry under its null; calling a procedure nonparametric does not remove all assumptions.
Reporting discrepancies to resolve before reproduction. Both the manuscript and repository text say17 posterior channels but enumerate16, apparently omittingP2. The alpha lateralization lists are asymmetric and includeP5 on both sides. Experiment2 claims six rotation magnitudes but lists seven, including26°, while Figure4 shows six. These are documentation inconsistencies, not proof that the executed analysis used the wrong channels or angles. Inspect the actual data labels and original scripts before silently correcting them.
Reproduction status and Amadeus implication. The earlier feasibility audit pins a later public code version and inventories49MAT files. A one-file download was prepared but interrupted before execution; no payload exists and no schema or empirical result was inspected. The owned download process was checked and none was running at session close. A focused Experiment1 reproduction remains an available next step, with a separate blind-orientation acquisition test needed for an extraction claim. The broader contribution is a concrete way to investigate access-dependent, item-level information beyond category evidence without assuming the measurement recovers an entire personal episode.