About RocSite.
Independent adversarial validation for clinical AI, built on two decades of enterprise architecture and regulated-industry experience. This is what we do, and why.
Why we exist.
Clinical AI is producing predictions that are treated as evidence, and most of them have never been adversarially tested by anyone with skin in the game. RocSite was built to do exactly that. The team brings two decades of enterprise architecture and regulated-industry experience to one problem: making medical AI safe enough to trust, and proving it.
How we work.
We pre-register. We lock the hypothesis before touching the data. We publish the result regardless of direction. Our methodology is designed to survive review, and our systems are designed to produce receipts, a verifiable artifact behind every claim.
What's running now.
RocSite, Inc. is a Delaware C-Corporation, established in 2022 when clinical AI validation became the focus. Bootstrapped. No outside capital. A small team that ships.
ICH Triage received an FDA Pre-Submission acceptance in January 2026 on the Class II SaMD pathway. Working ICU physicians have run the demo and exhausted their trial allocations in single sessions. AI Governor is at v1 with 21 active modalities: 19 deterministic clinical scoring specialists plus chest X-ray and head CT vision pipelines, purpose-built for the FDA-regulated AI vendors and institutions that need it. Caliper is live in production at RocSite Discovery: adversarial pharmacovigilance using FAERS disproportionality math, with a signed, public ledger of findings.
The methodology foundation is the sepsis falsification study, 286,510 ICU patients across MIMIC-IV and eICU-CRD, pre-registered before any data access, with the hypothesis locked. The methodology was reviewed by an established sepsis informatics researcher in connection with MIMIC credentialed access. The primary cohort returned a null result against our own hypothesis. We published it anyway.
That last sentence is the company. We pre-register. We lock the hypothesis. We publish the result regardless of direction. Everything else flows from that discipline.
How we operate.
We read every email ourselves. We anchor every claim on this site to a verifiable artifact, and if you find a claim that isn't, write to us and we'll correct it. Whatever you've heard about how startups operate, that isn't how this one operates.
We're not trying to look like anything other than what we are, a small technical team building governed, verifiable AI for medicine, with the discipline of people who would rather be right than impressive.
Anchored, not asserted.
The parts of our work that can be independently verified are below. Click through to confirm any of them yourself.