Preprint in preparation · not submitted · not peer reviewed

The paper

An LLM in control of an ensemble smoother: a pre-registered null, and why single-run controller comparisons in history matching measure the prior draw. Singh, Maiti and Patwardhan. Author order is not yet fixed.

Read it here

Embedded from public/paper/agentmatch.pdf. On a phone the HTML build is the better read: it reflows, and it is what arXiv’s own HTML view will show.

Two ways to read it, and one build note

The PDF is the typeset version, 15 pages, built with pdflatex from the source in paper/. The HTML beside it is the arXiv-style rendering, produced by LaTeXML from the same source. Both are generated from one main.tex, so they cannot say different things.

arXiv renders HTML through LaTeXML, and TikZ figures are the usual casualty. Ours were: on the paper as it first stood, both figures dropped and the bibliography failed as collateral, because the pgf binding errors aborted bibliography processing. The fix is in the source rather than in a workaround. Each figure wrapper switches on \iflatexml between the TikZ body and a pre-rendered image produced from the paper’s own preamble, and the pgf stack is not loaded at all under LaTeXML. The build that produced the HTML on this page is stamped in its own header: LaTeXML 0.8.8, Fri Sep 4 12:35:40 2026. One caveat we state rather than hide: arXiv’s converter is newer than the one we ran locally and carries its own bindings, so the arXiv preview still gets a look of its own.

A second note on the HTML, which is a site defect rather than a paper one. The generated file links its stylesheets with a path that resolves outside the web root when it is served, so it would render unstyled. The generated artifact is left exactly as LaTeXML wrote it — editing it would put the site and the generator out of step — and the server rewrites the request instead.

Every number in the paper is generated

No measurement is typed into the prose. Each one enters through a macro written by a script from an artifact in results/, which is why the prose, the tables and the figures cannot drift apart. Three rounds of review found the same defect each time — a number corrected in one place and left standing in another — and the macros are the only thing that has stopped it.

This site reads those same macro files. The headline figures below came out of paper/figures/f2-stats.tex when the page was served, not out of this page’s source.

Pairing paid: between-arm correlation +0.973, a pairing gain of 5.83×, within-pair sd 7.219 against arm standard deviations 28.77 and 30.69. The replication is reported beside the originals: same depth on 2 of 3 seeds, same schedule on 1, and on that one every scored metric reproduced to numerical precision. It is never pooled into k.

What still blocks submission

• 4 author notes remain in the source, at main.tex lines 49, 355, 917 and 934: author order and affiliations, the EOR scope call, the model versions and dates per period, and the author-contributions paragraph. The pre-send gate is make -C paper check — the Makefile is paper/Makefile, not one at the repository root, so plain make check fails for an unrelated reason. The target greps main.tex for the lines that carry those markers, with a threshold of zero, and exits non-zero while any remain; it builds main.pdf first, so it runs pdflatex before it counts.
• Three reviews have not happened: a scientific review, a reproducibility review from a fresh machine, and an adversarial review, ideally external. Nothing in this repository has been reviewed end to end by a human, and the adversarial agent passes that found forty-six defects in one afternoon — several of them in their own author’s edits — are a floor and not a ceiling.
• Two full texts are unobtained. The EAGE pitfalls paper, 10.3997/2214-4609.202539039, is the only open item that could force a rewrite rather than an edit, because its pitfalls section could pre-empt a claim. SPE-229413-MS is cited from its deposited abstract, and the paper says so.
• Open scope decisions, listed in docs/WORKPLAN.md: which comparison is Section 5.2’s primary result, whether Norne is in, whether to reimplement a baseline, author order, the localisation arm, the escalation suite, and what to do with the Vercel project whose premise this branch changes. The decision as written in the workplan is to disconnect it because the repository had nothing for Vercel to build; that is no longer true, so the open question is whether to keep the project and let it serve this site. The onboarding page carries it in full.
• The container has not been rebuilt from the patched dependency pin, and nothing is published as runnable until it is.
• The active-cell count is settled, and it is three numbers. The Data section states both the 35,939 cells OPM Flow simulates — the grid this site renders and the one the posterior lives on, read from public/data/manifest.json — and the 38,466 the deck declares in its own ACTNUM, and says why a reader who opens the deck and a reader who opens the simulator’s grid file get different answers. Neither is wrong; an earlier correction that called the deck’s figure a memory error was itself the over-correction, and was reverted.

The reproduction gate’s recorded verdict is GATE PASSED at 0% largest relative difference over 8 metrics, obtained on driver sha256 d1b858e2aa98… on 2026-08-25. The driver has been edited since this hash (guards added 2026-09-02/03). The analysis code paths it certifies are unchanged; a re-run of the gate against the current driver is listed in the WORKPLAN. It does not certify the current driver.

Author contributions and the use of AI systems

This paper studies language-model agents performing scientific work, and a language-model agent performed a substantial share of the work reported in it. The paper says so at length rather than in a footnote: which model, what it implemented, that it drafted the manuscript and two of the pre-registrations, that it ran the adversarial audit that found defects in its own earlier work — including a leaked answer key that invalidated an escalation result.

It also records a fact a reader should not have to discover: commit authorship in this repository does not distinguish human-written from agent-written work. The human authors are responsible for every claim, including any error originating in agent-produced work. No AI system is an author.