Every physical technology needs a data standard for safety.
How we think about making physical AI reconstructible, safer, and insurable: our white paper, writing from the team, and the published research we build on.
White paper · Working draft v0.8 · June 2026
Architecture of a physical AI trust & safety model.
Why physical AI breaks traditional insurance, and the architecture that fixes it: a measurable safety constitution, an engineering-derived risk prior, and PAIDS, the incident standard that is at once the underwriting engine and a cross-OEM safety knowledge base.
Articles
The black box for robots
Why physical AI needs a data recorder, and how a read-only side-car captures the incident window without ever touching the control loop.
8 min read →
Every household shift creates a new insurance category
Renters in the 1980s, devices in the 2010s, robots in the 2020s. The incumbents are always the last to notice the household has changed.
6 min read →
Out of the loop, by construction
How the RDR stays provably safe: no publisher on any command topic, a signed read-only policy, and an audit package your safety engineers can verify.
10 min read →
Published research we build on
From entropy to epiplexity: rethinking information for computationally bounded intelligence
Finzi, Qiu, Jiang, Izmailov, Kolter & Wilson (CMU / NYU). Why the learnable content of data, like a fleet's incident corpus, can't be valued by classical information theory alone.
PDF →
Bayesian safety validation for black-box systems
Moss, Kochenderfer, Gariel & Dubois (Stanford / Xwing). The certification math behind pricing rare events: estimating failure probability with a probabilistic surrogate.
PDF →
Sampling and estimation for the validation of safety-critical autonomous systems
Harrison Delecki (Stanford, Ph.D. dissertation, 2025). How to validate systems whose failures are too rare and too costly to observe directly.
PDF →