Reproducibility archive: A candidate-geometry principle that predicts hallucination risk, and physical hypothesis reasoning that abstains

Everything needed to recompute every number in the associated manuscript, on a machine that has never seen this work. The archive contains the experiments, each declaring in its own source what it writes, what it tests and which section of the paper uses it; the forward models that compute the physics the results are about, vendored with the path and hash they were taken from; the recordings the inherited figures rest on, including 443 archived robot plans as they came off the controller, the per-move table derived from them and the device-adapter scope table; the language-model transcripts of the open-world arm, every prompt and reply; and the checks, each of which exits non-zero when a claim stops matching its data. Each result file is stamped with the seeds and a hash of the code that produced it, so a file can be checked rather than trusted. The archive is tested by extracting it into an empty directory and running it there. Not included, for licence reasons rather than size: the two third-party driver corpora, 146 MB of other people's source. The archive records the repository and the exact commit for each, so the clone is reproducible without redistribution.