Lock Before You Compute: A Pre-Registration Workflow for Theoretical Physics Done with an AI Assistant (v2: lessons from an audit night)
Year: 2026
Type: preprint
Abstract: Status update, 20 November 2026. This version adds a one-page addendum (24346127_addendum.pdf) after an independent referee check: verified errata, what has changed since publication, and links to the newest records. The original files are unchanged. Newest records: SERAPIS v3.4 status update (doi:10.5281/zenodo.24396665); The Quark Thread (doi:10.5281/zenodo.24399408); Weave Control and the Evidence Atlas (doi:10.5281/zenodo.24399447). v2.1 (18 Nov 2026, refereed revision): Rule 3 corrected (12%, not 13%); rule 4's example is stated as a low-risk prediction guaranteed by a known identity; rule 6's claim withdrawn; prior art added (Nosek et al., Chambers, Klein & Roodman, Kerr). License (from this version): CC BY-NC-SA 4.0 under the AEGIS Community License — free for individuals, scientists and students; any business or for-profit use needs the AEGIS Commercial License (https://github.com/dannyeldridge54/AEGIS/blob/master/licensing/COMMERCIAL_LICENSE.md). Built by one independent researcher with the help of AI. v2 (18 Nov 2026). Adds an addendum with six rules from one audit night: (1) trace cross-checks to their inputs and void circular ones; (2) corrections are superseding records, also in machine memory; (3) evaluate a varying response where it is used; (4) write the prediction before the test (E6/E7, confirmed); (5) print small numbers raw; (6) a machine proposer needs a grader, and every deploy is checksum-verified. The v1 core is unchanged. Original files unchanged. Theoretical work can drift: a parameter is retuned after seeing a result, a failed check is quietly rerun, a remembered citation is wrong. We describe the workflow behind a series of inflation papers, developed by an independent researcher working with an AI assistant, that makes such drift visible. Every computation is preceded by a pre-registration file with its inputs, its pass criteria and a stated prediction, hashed and pushed to a public repository before the result exists; fixes go in a deviations file written before any rerun; every claim carries a status label; and predictions are frozen with public hashes and scored only on later data. We report what the workflow caught in one day of work (16 November 2026): 11 correction notes among 95 research notes, including a literature scaling used outside its regime that moved a key mass window by a factor of about 100, a pre-stated prediction that the locked run proved wrong, and a citation written from memory that stated the opposite of its source and had reached a published paper. The published archive reproduces all 22 key results from a clean unpack. Status of the claims. Counts and cases are taken directly from the public record (github.com/dannyeldridge54/aegis); they describe one intensive day and are not a rate. Single author, AI-assisted, no external referee.
Author: Eldridge, Danny
Language: English
Subfield: Library and Information Sciences
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7234183274
