For AI agent liability underwriters
How do you price AI agent liability when there is no actuarial history? Insurance needs decades of claim data to set premiums — autonomous AI agents are barely five years old. Sgraal's fleet generates millions of real preflight decisions per month; we are structuring that into the first AI risk dataset.
Honest disclosure: this is on our long-term roadmap, not yet built. No demo calls, no follow-ups until there's something tangible to show. Email above just adds you to a low-volume list (~2 emails/year max).
Historical incident-rate distributions from the Sgraal fleet aggregate. Anonymised, differentially-noised, structured for actuarial modelling rather than raw access.
Cox-style proportional hazards models of memory state survival — how long until a given agent's memory crosses the BLOCK threshold under typical and adversarial load.
A trustworthiness composite (behavioural profile, counterfactual block rate, healing convergence) intended to be referenced in per-agent or per-fleet liability policies.
A long-horizon financial primitive: action-level mikro-bonds priced from agent risk score and outcome data. Far-future concept; we mention it to be honest about the direction, not to promise it.
The concept items above build on top of these production features.
Already runs inside the production scoring engine. The actuarial maths is in place; the per-customer dataset productisation is the new part.
See the discipline list →Fleet-wide "would BLOCK have been correct" statistics are already measured per tenant and per domain.
Read the whitepaper draft →Decision boundaries already adapt to each tenant's traffic after an initial calibration window. The underwriting-grade version of this signal is what the dataset license would expose.
See platform features →If AI liability underwriting or risk consulting is your domain, leave a note. We email back when there's something to show, not before.