Preregistration freeze: source-confound controls for a trilingual line-drawing retrieval benchmark

This archive fixes the content of a preregistration outside the repository that holds it. The study it belongs to reports negative results: a set of retrieval conditions that do not separate from chance once source and page type are held constant. Negative results depend on the hypotheses having been fixed before the runs, so the ordering has to be checkable. Within the repository it already is, because every run artifact records a commit hash that could not have been known before that commit existed. But that repository is one the authors control. What this archive does and does not establish. It establishes that the preregistration had exactly this content no later than the deposit date. It does not, by itself, establish the freeze date: that rests on the commit object and on the hashes recorded in the run artifacts. Stating this narrowly is deliberate, because an external deposit is not a timestamp for events that preceded it. Contents. The complete tree of commit efa455c1405ac9a46ccc7e198d6f89d4d91557dc, dated 2026-08-27T18:42:27+01:00, minus images: 63 files, 2.36 MiB. The 1,699 excluded image files are third-party material not authored by this project, and every one is listed in MANIFEST.json with its sha256, so the exclusion is verifiable rather than a gap. The archive contains no analysis script and no result. That is what a preregistration means: this commit predates every run. MANIFEST.json carries the commit hash, the commit date, and a sha256 for every file both included and excluded.