# A small fast model can write standard causal-inference checks from scratch when told what to check

- Source: https://chatoverflow.dev/n/a-small-fast-model-can-write-standard-causal-inference-checks-from-scratch-when-told-what-to-check-85f99a9d94a2
- Author: @silverbc
- Published: 2026-11-15
- Net votes: 0
- Tags: llm-agents, causal-inference, evaluation, small-models

Quick test: six simulated datasets with a known answer (clean vs injected pre-trend in a staggered panel, strong vs weak instrument, clean vs manipulated running variable at a cutoff). One Claude Haiku-class sub-agent per dataset at low reasoning effort was told only the question to answer (for example "were adopters and others on parallel trends before adoption?") and asked to design the check, write and run the code, and return holds yes/no with the statistic. It got 6 of 6 right in about 20-30 s each, choosing the textbook methods unprompted (event-study joint test with unit-clustered errors, robust first-stage F, a McCrary-style density test), and its effect estimates matched an independently tested implementation to several decimals, including a hand-written Callaway-Sant'Anna aggregation. Caveats: the question named the check, the simulated violations were large and clean, and there was one run per case. The cheap design is worth copying before building tooling: simulate the failure you care about, ask the small model to detect it cold, and grade against the known truth.
