Do cats rot your teeth? Preregistration, code, outputs, process logs and transcripts of a study produced by an AI agent (NHANES 2005–2006)

This record holds the preregistration, code, outputs, process logs and session transcripts of a study that is also a meta-experiment on research produced almost entirely by an AI agent (Claude, model claude-opus-5-5, Anthropic, run in Claude Code) under the direction of one researcher.

The substantive study asks a deliberately implausible question: is living with a cat associated with untreated dental decay among US adults? It uses public data from NHANES 2005–2006 with survey-weighted modified Poisson regression, prespecified confounders, sensitivity analyses and equivalence testing. The preregistered primary result was inconclusive (prevalence ratio 1.12, 95% CI 0.93 to 1.36). The meta-experiment measures what producing the study took: time, human interventions, agent errors and the share of code and text written by the agent. The accompanying preprint, "Do cats rot your teeth? A by-the-book paper written by an AI agent, and what it means for scientific communication", shows how four of the 45 preregistered estimates could support a positive paper without breaking the registration, and argues that verification, not writing, is now the bottleneck of scientific communication.

Version 1 (doi:10.5281/zenodo.23435708) is the frozen preregistration, deposited before any exposure–outcome association was computed and later registered on OSF (doi:10.17605/OSF.IO/RVH7J). Its files remain available in version 1.
Git tag: prereg-v1.3
Commit: 4a7438cf1e489ec46a998b4b73c4f1c469485de4 (2026-10-17 08:11:38 UTC−3)
SHA-256 of prereg_v1.3.md: b3a8886a4f8154d0e263aaacb68650e012c0f6486d30ffaf850e0cc6f8e197ba

Version 2 (this version) adds the project repository at the preprint, the redacted session transcripts (including the agent's reasoning) and the preprint PDF.
Git tag: preprint-v1
Commit: 0929be774473da2614b094a164766e1ebda26609 (2026-10-18 16:41:01 UTC−3)
SHA-256 of each file: CHECKSUMS.txt. Correction (metadata edit, 18 October 2026): the code archive deposited here, paper-cats_preprint-v1.zip, has SHA-256 090329d96bd4ce918216c6882cd76a55e23ebe130688047358e402efa072373a. CHECKSUMS.txt lists a later rebuild of the same archive (bc12a52666b4168b44e93695e5a26b188d35d643ea676a1c99a0569fb7b11867) with identical contents (the same 183 files, byte for byte) and different zip timestamps. The other files' checksums match.

Not included: full texts of cited articles (copyright) and the raw NHANES files (public; re-downloaded by scripts/01_download.py, with SHA-256 checksums in the repository). The transcripts are redacted for personal data and for article texts returned by tools; see README.md.

Status: process figures are as of 18 October 2026. Logging continues through peer review until publication or the author's recorded decision to stop; new versions of this record are deposited when the manuscript changes materially and at the end.

Licenses: code in scripts/ under the MIT License; all other content under CC BY 4.0.

Author: Diego Halabi (Universidad Austral de Chile). The agent wrote the code and drafted the protocol and manuscript under the author's direction; the author made every methodological decision from the options offered by the agent and wrote the conclusions. The agent is not an author.

The version posted on medRxiv (git tag preprint-v1.2) has the same content as the PDF in this record except for one corrected statement (the abstract and Methods said the agent drafted all the text; the author wrote the conclusions and most of the title) and a typeset layout. Process figures are unchanged (cut-off 18 October 2026).

Correction (metadata edit, 22 October 2026): the preprint's Methods say that the AI agent "formulated the question", and the repository README said that "the agent formulated". This is wrong. The research question (whether living with cats is associated with worse oral health), the thesis, the data source (NHANES 2005–2006) and the exposure variables were posed by the author in the first message of the project; the agent operationalised the question by offering options (population, outcomes, covariates, analysis), from which the author chose. The preprint's own "Use of AI" statement says so ("The author set the question and the project rules"), so the preprint contradicts itself. The author found the error and logged it as ER45 in log/error_register.md; later versions corrected it.

Preprint status (22 October 2026): the preprint was submitted to medRxiv on 18 October 2026 and declined at screening as out of scope ("Simple automated/computational analyses of public data ... are generally not sufficient"). This record is therefore the public version of the preprint; there is no medRxiv DOI.