The accumulation problem
The denominator problem is about pricing one risk. The accumulation problem is about surviving a portfolio of them — and it is the part of the thesis that speaks to reinsurers rather than underwriters.
One update, every deployment
Section titled “One update, every deployment”A foundation-model update by a major AI provider can change the behavioral profile of every deployment built on that model, simultaneously. If ten enterprises covered by different syndicates deploy systems on the same foundation model, a single upstream change creates a correlated loss event across the portfolio. No syndicate currently has visibility into that correlation. GROUNDED · Brochure
Trace the dependency chain to see why the blindness is structural:
Each syndicate sees its own insureds. None sees the shared upstream dependency, because nothing in a policy application captures “which foundation model, which version, updated when.” The correlation lives in a layer the market has never instrumented. GROUNDED · Brochure
Why this is the reinsurer’s nightmare
Section titled “Why this is the reinsurer’s nightmare”Volume 1 taught that reinsurance and capacity exist to absorb exactly this — correlated, portfolio-level loss. But a reinsurer can only model accumulation it can see: hurricanes have tracks, cyber has named vulnerabilities and vendor concentrations. AI accumulation today has neither a map nor a census. The brochure: reinsurers “cannot model accumulation because no one can identify how many policies contain correlated AI behavioral risk.” GROUNDED · Brochure
| Peril | Correlation carrier | Can the market see it? |
|---|---|---|
| Hurricane | Geography | Yes — mapped, modeled since Andrew |
| Cyber | Shared software/vendors (e.g. a compromised library) | Partially — vendor questionnaires, scanning |
| Generative AI | Shared foundation model + version | No — not captured anywhere today |
What evidence at the boundary changes
Section titled “What evidence at the boundary changes”The countermeasure follows directly from the instrumentation: if every governed inference is attested against a declared baseline, then version change and behavioral drift surface as events, in real time, independently of the vendor’s own reporting (mechanics in baselines and drift). Aggregate those events across deployments and the same upstream model update shows up as a synchronized signal across the book — the correlation becomes observable for the first time. GROUNDED · Brochure
The draft application makes this a program deliverable: an accumulation-facility demonstration — showing at least one reinsurer how Control Platform data aggregated across multiple insured deployments provides “portfolio-level AI behavioral-correlation visibility.” That is a proposed ten-week-cohort deliverable in a draft, not a shipped capability — hedge it accordingly when selling. GROUNDED · Brochure
The long arc
Section titled “The long arc”Accumulation visibility is also the opening move of the longest play in the deck: a loss model built on an attested inference population is the prerequisite for transferring AI systemic risk to capital markets, the way catastrophe bonds transferred hurricane risk after Hurricane Andrew forced the industry to build modelling infrastructure. That arc — explicitly a 5–10-year horizon — is AI liability ILS. For now, hold the shape of the thesis: define the unit (the inference), count it independently, control it at the boundary, and both pricing and accumulation management become possible. GROUNDED · Brochure
The next section descends from thesis to machinery: what actually sits at the inference boundary, and what it does on every request.
Drill this page →20 bank questions stand behind what you just read. Check it while it’s warm.