Professor of economics at Charles University, Prague. I work on meta-analysis, publication bias, and p-hacking, mostly with Zuzana Iršová, and have published in journals including the Review of Economics and Statistics, the Journal of the European Economic Association, and Nature Communications. Affiliated with CEPR and the Stanford METRICS center, and a former advisor to the Board of the Czech National Bank.
Our work on meta-analysis is at meta-analysis.cz: 71 papers in full, 46 datasets with 58,405 estimates, and the estimation code, all under CC BY 4.0, AI training included. The site is also a public repo, built to be read by machines as well as people. The papers have been cited more than 10,000 times on Google Scholar, including by five Nobel laureates, and more than twenty central banks, the IMF, the research service of the US Congress, and the US Environmental Protection Agency draw on them. The impact page shows where and how. On methods, MAIVE (Nature Communications, 2025) corrects meta-analysis for spurious precision, and easymeta.org runs it in the browser.
I also build experimental AI tools that bring adversarial review to empirical papers, grant proposals, and referee reports. The most ambitious is a Claude-only workshop (CRUCIBLE) that argues a paper to the sentence and then, optionally, rebuilds it. It grew out of a manual, human-in-the-loop audit protocol (ChatGPT, Gemini, Grok, Claude) and its automation for Claude Code and Codex. Prof. Bob Reed (University of Canterbury / MAER-Net) has independently endorsed the protocol. The tools surface criticism for a human to weigh.
We tested them against the simplest alternative, expecting them to win. In a pre-registered, identity-masked experiment, the authors of 44 meta-analyses ranked AI reports on their own papers by usefulness, each report held to the same length and template. A single careful pass by a frontier model came out ahead of both mad-research and a deliberately light desk-review configuration of CRUCIBLE, though one of three AI judges, Gemini, ranked CRUCIBLE first. The fuller workshop modes and the rebuild stage were outside the test, as was the manual protocol. Havránek and Iršová (2026), arXiv:2607.14713; pre-registration, data, and code at meta-analysis.cz/debate. For feedback on a finished paper alone, start with a careful single pass. The tools below earn their keep on quote-grounded comments at scale, a cross-model second opinion, and the rebuild.
For students at Charles University there are two web apps with nothing to install: an unofficial check of an application to the university's grant agency (GA UK) before submission, and feedback on an economics thesis at the Institute of Economic Studies before the seminar.
| If you want to… | Use | What it is |
|---|---|---|
| Get data, code, and full texts for our meta-analyses | meta-analysis.cz | One folder per paper with the full text, the data (CSV and Parquet, with column-level codebooks), and the estimation code, plus a pooled estimate-level table. CC BY 4.0. |
| Correct a meta-analysis for publication bias, p-hacking, and spurious precision | MAIVE · app | The estimator from our Nature Communications paper, as an R package on CRAN and as easymeta.org, which runs it in the browser with no code. Petr Čala wrote the app and most of the package (app code). |
| Workshop and (optionally) rebuild a whole paper | paper-workshop | CRUCIBLE: a Claude-only fleet of rival-tradition referees argues the paper to the sentence, every comment grounded in an exact quote and no acceptance-probability scores. On request it rebuilds: a tracked redline, your own code re-run so the numbers are real, and a replication package. |
| Run a manual, human-in-the-loop adversarial audit of a high-stakes paper | research-audit-duel-protocol · app | The protocol the automations grew out of: the Duel + MAD workflows across ChatGPT, Gemini, Grok, and Claude, with a worked example. The page walks you through a run and hands you each prompt, ready to paste. Independently endorsed, and the methodology to cite. |
| Automate that audit across Claude + Codex, without writing code | mad-research | Three Claude Code skills: one-shot Codex calls, collaborative build with cross-review, and a three-stream adversarial audit that ends in a memo. |
| Triage an ERC Starting/Consolidator draft before peer review | erc-ai-feedback · app | Unofficial, though I served on the ERC Advanced Grant economics panel in 2020 and sit on the expert group that supports Czech ERC applicants. Paste one prompt into a chatbot, then your draft, and it flags the routine problems a busy panel member would see, such as vague hypotheses or weak panel fit, before workshop time is spent on them. It does not write proposal content or replace human review. |
| Check a GA UK application before submitting it | gauk-ai-feedback · app | Unofficial, in Czech and English: formal problems first, then whether the application says what the project will add. The attachments are checked in the browser without being uploaded, and the tool never writes the application. I sit on the GA UK Grant Council (see the disclosure). |
| Get feedback on an economics thesis before the seminar | ies-thesis-feedback · app | For bachelor's and master's students at the Institute of Economic Studies: send a chatbot anything from an idea to a finished thesis and get it read the way your supervisor and opponent will. |
| Curious what kids can build? | race | A two-player browser racing game our kids (ages 8–12) made with Claude Code, set in our town of Litomyšl. We helped them publish it. Since then they have made Země Ňuňísků, a real-time strategy game in the spirit of Age of Empires where you pick a civilisation, gather resources, build, and fight bots or friends online (I helped with it for a few days in summer 2026), and a series of teaching films, from ancient Rome to the periodic table, with AI-made images, music, and narration. |
The workshop (CRUCIBLE): Havránek & Iršová (2026), doi:10.5281/zenodo.20828996. The audit protocol it grew out of: Iršová & Havránek (2026), doi:10.5281/zenodo.19105954. The experiment: Havránek & Iršová (2026), arXiv:2607.14713. The datasets: doi:10.5281/zenodo.21773678, together with the paper whose data you use. Most tool repos ship a CITATION.cff, so GitHub's "Cite this repository" works there.
- Website: tomashavranek.cz
- Google Scholar: Tomáš Havránek
- ORCID: 0000-0002-3158-2539
Getting started: erc-ai-feedback, gauk-ai-feedback, ies-thesis-feedback, and research-audit-duel-protocol need nothing installed. All four are web pages that hand you a prompt to paste into any chatbot, and only the first three have optional agent routes that need setup. CRUCIBLE and mad-research are Claude Code skills and need a Claude subscription, and mad-research also calls Codex. Claude Code runs in the Claude desktop app or in a terminal. Each repo's README has the steps.
A note on confidentiality: an AI review sends your text to the provider of the chatbot or agent you use. Local checks, such as the GA UK app's file check, send nothing. Each README says what goes where, so read it before sending anything embargoed, under double-blind review, or otherwise non-public.