Pre-registered falsification protocol and results

For hedge fund research reviewers, this report examines checks for selective reporting when a registered study is closed. At Axio, the proposed use is review of research records before their conclusions enter an evaluation; no product integration or saving in review time has been measured. This report makes no claim of originality. Checking whether a study followed its pre-registration requires comparing the report with the plan. This report tests one method for that comparison. A research record is kept in a fixed JSON format (registration, hash commitment, event log, closure), and seven mechanical rejection rules, R1 to R7, each target one selective-reporting defect taken from Arnott, Harvey and Markowitz (2019) and Harvey and Liu (2015). The comparison arm gives the same information, rendered as plain text, to a general-purpose language model (Claude Sonnet 5.5) with one fixed instruction. The test plan, its thresholds and its decision order were frozen, with a local SHA-256 hash, before any code, record or label existed. For each of two cohorts a separate model session generated 33 synthetic records, 12 valid and 21 with one injected defect, without seeing the rules, and labels were locked before either arm ran. In the first (in-sample) cohort the rules flagged 21 of 21 defective records and rejected 0 of 12 valid ones. The language model also flagged 21 of 21 but rejected 6 of 12 valid records, which disqualifies it under the plan. In the second (out-of-sample) cohort the rules again scored 21 of 21 with 0 of 12 rejected, and the language model scored 17 of 21 with 0 of 12. The gap of 4 is below the pre-registered margin of 5. The pre-registered verdict for both cohorts is therefore "compared, threshold not met", and the report makes no claim that the rules add detection. A reduced rule set (R1, R2 and R7 only) missed 12 of 21 defects in both cohorts and is refuted. No confidence intervals are reported, because the plan forbids them: the records are constructed counterexamples, not a sample from any population. Deposit version: Working paper v1. Author: Peilun Cai. Affiliation: .The manuscript is an English typeset presentation of the same experiment; no new experiment or result was introduced. PDF SHA-256: 7cdc7abe30104c8629dc37fd829cc200bb679a03e6b92469d50dd96c257aea06. Replication archive SHA-256: c51be0f20a58d65775920162677b9fdcc91d3c37b0d306bf708013fd3e07d10d.Independent review of this revision remains pending. Public pre-result timestamp receipts: 0. Local hashes and clocks do not establish public pre-registration. The registered secondary outcome about reasons was not computed; missing is not zero.The replication archive contains English presentation copies, original hashes, a current source mapping and a public numeric checker. Original-language records available on request, verifiable against listed SHA-256. The original historical review is withheld; its scope and findings are represented by an English reproduction summary. The source and PDF are supplied separately.Rights, funding, conflicts of interest and author review remain for the human author to confirm. Declaration of generative AI and AI-assisted technologies.During the preparation of this work the author used Claude (Anthropic) and Codex (OpenAI) to write analysis code and draft text. After using these tools, the author reviewed and edited the content as needed, verified every number and citation against the replication package, and takes full responsibility for the content of the publication.