Bibliographic search archive for "Predicting Representations, Modeling Worlds, Recovering Causes: A Critical Review of JEPA, World Models, and Causal Representation Learning"

This archive supports the methodology section of a critical review of JEPA, world models, and causal representation learning. It contains the material needed to verify every count reported in the review's identification-and-screening flow diagram, and to re-run the retrieval. The search reported in the review is an audit-and-extension arm, not the path by which the review's corpus was assembled. It was executed after the first version of the review was complete, across OpenAlex, Scopus, and Web of Science, in six query families corresponding to the lineages reviewed: state-abstraction and sufficiency theory, self-supervised and joint-embedding prediction, model-based reinforcement learning, causal representation learning, object-centric learning, and embodied world modeling. Contents: the six family queries and per-database manifests; the retrieval, screening, and metadata-verification scripts as executed; the funnel report; the 264-record candidate pool with per-record verification status; the per-record screening decision log; the recall-audit target set; the index status of every record the search missed; and the full retrieved records for the OpenAlex and arXiv arms. Scopus and Web of Science records are not redistributed. Their raw form includes abstracts, which the respective terms of use do not permit redistributing. Those two arms are represented instead by an identifier-only projection generated by this project, which is sufficient to reproduce the identification totals, the pre-screening removals, the duplicate merge, the unique-record total, and the database-overlap figures. A verification script included in the archive recomputes those counts from the deposited material alone, by importing the deposited screening code rather than paraphrasing it. The rule-based scoring step reads titles and abstracts and therefore cannot be re-run for records appearing only in Scopus or only in Web of Science; the queries are deposited so that a reader with database access can re-fetch those arms and re-execute the pipeline end to end. Headline counts: 35,781 records identified; 871 removed before screening; 13,802 duplicates merged; 21,108 unique records screened; 3,890 surviving rule-based screening; 264 candidates assessed and metadata-verified. See README.md for the full contents listing, the licensing and provenance breakdown, and a note on what can and cannot be recomputed.