CausalSmith: An Agentic Pipeline Produces Lean 4-Verified Econometrics Papers with GPT-5.6 and Claude
daveholtz · x · 2026-09-15
Vasilis Syrgkanis's team at Yale, with student Jiyuan Tan, released CausalSmith, a fully automated agentic pipeline for theoretical research in causal inference.
- Every formal statement, theorem, assumption, and lemma in its working papers is machine-verified in Lean 4, built on the Causalean library with natural-language translations and review status.
- Division of labor: OpenAI's GPT-5.6 Sol does the main math and Lean formalization, Anthropic's Claude Opus 4.8 handles planning and Lean code review, and OpenAI's GPT-5.5 drafts the write-ups.
- Featured paper establishes sharp minimax MSE rates (1/n + d²/(n²(log n)²)) for average treatment effects with discrete confounding under fixed overlap, scored 8/10 by an AI reviewer.
A rare "AI scientist" attempt where every mathematical claim is machine-checkable rather than self-reported by the model.
More from coding & agent
- The Best IDE for Agentic AI May Not Be an IDE at All — rseroter · 2026-09-16
- Harness Engineering: building the environment that makes coding agents reliable — Pavan_Belagatti · 2026-09-16
- Flue 2.0.7 ships: context-overflow responses now survive compaction — ritakozlov · 2026-09-16
- Watching Codex build 3D game assets remotely from a phone — chongdashu · 2026-09-16
- Lobsters' broken Anubis lockout solved by Codex, which bypassed it to message moderators — mitsuhiko · 2026-09-16
- Three engineers livestream building a company in 3 days with Grok Bot — EricBuess · 2026-09-16