Integrated and separated event representations

Status: partial reading, stopped at the user's request. Do not count as a completed full-paper reading or a completed methodological audit.

Zhenghao Liu, Mikael Johansson and Inês Bramão. Episodic events are flexibly encoded in both integrated and separated neural representations. Nature Communications 17, 752. Published article, publisher PDF, PMC full text.

Retained PDF, extracted text, supplement, reporting summary, peer-review file, provenance manifest.

Scope completed

The retained 15-page PDF is the final publisher layout, not an unedited reference.pdf. The introduction, main results, discussion, figure captions and core methods were examined in extracted text. Bibliographic backmatter, final availability details and a complete visual check of figures/equations were not finished. All six main figures require visual verification before calling the reading complete. The 14-page supplement, 3-page reporting summary and 23-page peer-review file were successfully downloaded and parsed, but not read. The reporting-summary PDF has no useful extractable text; its empty derived text was removed. No OCR or further downloads were attempted after the stop instruction.

Provisional factual core — from the main paper only

Participants learned repeatedly presented animated AB and BC events sharing a character, alongside unrelated XY events. Of 41 recruits, 36 entered the main analysis; one ceiling performer and four EEG/programming exclusions were removed. Later tests measured direct and indirect AC associations and whether characters had appeared together. Related-event EEG similarity during presentation of C was associated with later AC success; at another interval, greater similarity accompanied worse source memory. Later dissimilarity during B accompanied better source memory. These positive results support temporally differentiated neural relationships to inference and source judgments. Main analyses omitted the first BC/XY exposure; each movie appeared five times. Neural–behavior analyses further restricted participants to those below or at 90% accuracy, leaving 30 for association and 28 for source analyses. Their regressions treated EEG similarity as the dependent variable and later accuracy as a predictor. They were associations with subsequent performance, not held-out forecasts. The main text reports no systematic AC-retrieval effect on surprise context/clothing-detail memory; its supporting supplement remains unread.

Provisional audit and competing interpretations

These are original evaluation points, not additional established findings or a completed replication.

• Useful controls: comparing corresponding AB–BC pairs against noncorresponding AB–BC pairs improves on a contrast against wholly novel XY events alone. Counterbalancing which second-phase movies count as BC or XY reduces fixed movie-content confounding. These controls support a relation-specific signal, while leaving retrieval, anticipation and stored integration as competing explanations.
• Feature selection: the main text selects informative EEG features using within- versus between-AB-movie similarity, without directly selecting on later behavioral labels. That is stronger than choosing features for their behavioral correlation. It nevertheless uses the same participants and AB recordings to construct the later comparison; independent generalization and complete statistical independence have not been established by this audit.
• Repeated exposures and visible features: the principal analysis follows prior BC encounters, so C can already cue its partner and associated AB material. This is not a clean measurement of the first formation of a novel AC representation. The shared B is itself visible at the later interval; comparisons need to distinguish perceptual, familiarity and mnemonic contributions.
• Source-response contract: the source question concerns whether a character pair appeared together. Before interpreting it as preserved episodic content, verify which pair was shown after an incorrect associative response and how correctness was scored. The present reading did not resolve that contract. The main text also reclassifies correct responses slower than 10 seconds or marked as guesses; inspect exactly which endpoints inherit that rule. Fast confident correctness cannot by itself distinguish stored inference from rapid retrieval-based reasoning.
• Dependence and selection: repeated movie exposures share a later test outcome. Whether the Bayesian neural–behavior model handles participant/movie nesting adequately remains unverified; coefficient standardization is not by itself a nesting model. Outcome-based ceiling exclusions restrict the analyzed population. A significant coefficient at one interval and a nonsignificant coefficient elsewhere do not establish a difference between intervals or functions.
• Coexistence versus mechanism: the coexistence of temporally distinct similarities and dissimilarities is useful positive evidence. Neither a null correlation between them nor scalp EEG establishes independent biological mechanisms, hippocampal localization, or two separately stored representations. The reported association of one similarity interval with worse source performance also prevents a blanket claim that integration is cost-free.

Original implication for functional extraction

A recipient architecture could preserve episode provenance while allowing relational retrieval across episodes; this paper motivates testing that combination rather than selecting a unique storage architecture. A fair test would compare stored integration against retrieval-time inference with matched inputs, test association and source judgments separately, and include genuinely new event combinations. Neural similarity is not an extracted personal episode or a demonstration that these representations can be transferred to a recipient.

Availability scout and unfinished work

Zenodo record 17612987 advertises preprocessed EEG, behavioral data and code. Its inspected API metadata lists one Open_Data.zip, 32,031,490,629 bytes; the archive was not downloaded, and its internal schema/code was not inspected or executed. The main text says raw EEG is planned for January 2028. Advertised availability is not a verified runnable reproduction.

Outstanding work is limited but consequential: read all supplementary notes and reporting summary; inspect figures; verify source scoring, trial/participant nesting, ceiling exclusions, temporal/multiple-comparison inference, and extraction of code/data from the large package if a later reproduction is chosen. No empirical reanalysis or transfer experiment was run.