Tianzhu Qin

8/ StatsClaw is open source. We build a /contribute command that can summarize your lessons for the community. Come build with us! statsclaw.ai

7/ It's not magic. Quality scales with how much you engage: discuss the plan, review the comprehension artifact, help design tests. Dense math can trip it up. But for package maintenance and new features, it changes the game.

6/ In another, the tester discovered a property that causes inconvergence we hadn't considered: balanced-panel geometry makes certain FE augmentations algebraically degenerate. The system didn't just verify — it discovered.

5/ We used it on our own R packages (panelView, interflex, fect). In one case, the reviewer caught 6 bugs that passed all 34 tests — including 2 that silently produced wrong standard errors.

4/ Here's a demo. We gave it a 4-page PDF with three probit estimators (MLE, Gibbs, MH) and one prompt: "Build the R package from this PDF. Run Monte Carlo. Ship it." What came back: a complete R package with C++/Armadillo backends, 3 estimators, a full test suite, and Monte Carlo results — all verified against R's glm().

3/ StatsClaw enforces information barriers. A planner reads your math and produces three independent specs — one for the builder, one for the simulator, one for the tester. None of them can see each other's instructions. The builder doesn't know the ground-truth parameters. The simulator doesn't know how the algorithm works. The tester just checks: does the implementation recover ground truth? A bug that survives must fool all three independently.

2/ Key idea: verification with information barriers. The problem: when an LLM writes both code and tests from the same information, the agents can cheat or find workaround then fixing the root cause. If it misunderstands your estimator, it embeds the same mistake in the tests. Everything passes. The implementation is wrong.

1/ Happy to release StatsClaw — an open-source multi-agent workflow for building statistical software with AI. w/ @Maple_Optboy Site: statsclaw.ai Paper: bit.ly/statsclaw