Articles

By Mark Wasuwanich · October 15, 2026

We ran the same audit categories repeatedly on two independently built AI agents. One category moved by 18.5 to 25 points on a 1-100 scale across runs, well past the tolerance we set, while several others held stable. Here is what that means for evaluation evidence in AI-agent underwriting submissions.

By Mark Wasuwanich · October 14, 2026

Pure premium is supposed to equal frequency times severity. We measured how often multi-task pricing models violate that identity across nine model families on two public motor datasets, and found a simple fix that removes the gap at no cost to accuracy.

By Mark Wasuwanich · September 5, 2026

A read through what people are actually building in and around insurance pricing right now, and what it points to for the model that eventually replaces the annual number.

By Mark Wasuwanich · August 29, 2026

We tested whether risk features an LLM extracts from 10-K text improve an accounting-risk model trained on financial ratios alone. The lift shows up at the tail, right where underwriting decisions get made. Here's the result and where we're taking it next.