ThetaDriven · Risk Brief

The accumulation no cat model sees

You are writing AI liability as if each policy fails on its own. They don’t. Every insured’s AI leans on the same three or four frontier models — so the book is one correlated position wearing the costume of a diversified one.

Illustrative — one shared event, simultaneous claims
sharedmodel⚡ version bump / jailbreak / API changeclaims,all at oncea “diversified” book…
The illusion
Proxy pricing (revenue, industry class, IT questionnaire) assumes independent, idiosyncratic failures. Law of large numbers holds only if that’s true.
The reality
One upstream event — a model version bump, a shared jailbreak, a prompt-injection vector, an API change — correlates the entire book. It fails together, not one policy at a time.
Why there’s no cat model
Cat models need history. AI drift has no loss surface yet, and it’s adversarial and non-stationary — you cannot backtest an accumulation you’ve never observed.

The number you don’t compute today

What percent of your automated decisions share a single model dependency? That figure is your true PML concentration — and right now nothing in the stack produces it. A signed drift receipt per run does: it records where each agent’s work landed on a fixed 144-cell lattice, recomputes byte-identical for any auditor, and fires before a claim exists. It is the forward sensor that builds the accumulation dataset going forward — the one you can’t reconstruct from a past that didn’t log it.

The honest part — and the ask

A drift receipt is a countable event, not yet a loss-linked one. Whether the drift we measure predicts your correlated losses is an empirical question — and only a house with a real loss book can answer it. That’s the proposal: a joint validity study mapping our leading drift signals against your actual incident record. Not a product sale. A measurement partnership on the one exposure that could end a book.

npx thetacog-mcp attest-demo
Recompute the receipt yourself — ten seconds, your hardware, no model in the loop.
Then: elias@thetadriven.com · a 20-minute conversation on the study.

The lattice diagram is illustrative. The measurement is open source (MIT); the financialization is licensed. See /brokers · /agent-year · ThetaDriven · US Patent App. 19/637,714.