Research
Most vendor research exists to be cited, not read.
These are working documents. Each states its method, its sources and the point at which it stops being able to support a conclusion.
The accountability gap in insurance AI
[V]What the Annex III obligations actually require of a decision, and what the current vendor field publishes about its own logging.
Deliberation as an engineering problem
[D]Why a panel with an adversarial seat and enforced evidence separation outperforms both a single model and a naive multi-agent debate — and where that stops being true.
What a computed gate can and cannot prove
[D]An agreement predicate between two heterogeneous models is not a proof. This sets out precisely what it does establish.
Catastrophe physics versus a catastrophe model
[P]What our peril simulation does today, and the financial chain it does not yet have. Written so an exposure manager can tell the difference.
The catastrophe piece is marked as a projection because the financial chain it describes is designed and not built. It sets out what our peril simulation does today and what it would still need to be a catastrophe model — read it as a gap analysis, not a capability claim.
Every claim in these is tagged the way the rest of this site is tagged: verified with two independent sources, observed in code we can show you, projected, or an internal method attributed as ours. Ask for the underlying run and you get the seed with it.