eLife Claim Trees
What a paper claims, where each claim came from, and whether anyone checked
A paper is written to convince a person reading it once. Its claims are embedded in prose, bundled into figures that carry several results at a time, and severed from the code and data that produced them. This project takes 10 eLife neuroscience papers apart into their individual claims — 289 of them, with 916 typed relations saying how each depends on the others — and links every claim to the panel, the data and the analysis behind it.
The unit is the panel: one analysis producing one result. Figures are too coarse, since a single panel usually carries several findings its caption runs together. A p-value is too fine to mean anything alone. The panel is the level at which reproducibility actually operates.
What is here
What this does not show
No person has checked any of it.
Every claim was found by a language model, reconciled by another, and related to its neighbours by a third. Approval is recorded per version, against the output a person actually read — and it stands at 2 of 310. Read this as a draft annotation layer, not as adjudicated output.
The corpus has a jagged edge.
Every paper goes through the same 31 layers, and no two have been through the same ones. Gadeke is furthest at 30 of 31; 140 cells across the corpus have never been run. The pipeline page says which, and each empty one says what would fill it.
Reproduced is not replicated.
Where a claim is marked reproduced, the authors' own deposited code was re-run against their own data. Whether an independent study found the same thing is a different question, and this corpus has never asked it — replication is declared and empty for every paper.
How it is made
A layer is a question asked of a paper. It reads declared inputs, produces a version, and records the content hash of everything it read — so whether a result is still current is asked of the files rather than of anyone's memory. The pipeline asks 31 of them of each paper, and 5 once of the whole corpus — the schema and vocabulary decisions every paper inherits. A paper's page carries its own instance of the 31.