Governance and audit infrastructure. Not a clinical-decision-support device. Not used on real patients. Inputs are synthetic, public-dataset, or de-identified.
For governance committees, CMIOs, and research teams

Evaluate clinical AI outside the live clinical workflow.

Your committee needs to know what a clinical-AI tool is actually doing, before procurement, or after deployment, without putting it near a patient. RocSite compares a model’s output to a locked, published clinical instrument on synthetic, public, or de-identified data, and signs the result. What you receive is an independent, off-patient audit record you can hand to a committee or a regulator.

Ed25519-signed evaluations  ·  OSF pre-registered  ·  MIMIC-IV + eICU-CRD validated  ·  Never a third reader in the worklist

See AI Governor → Request access
6
Confirmed findings, replicated on two datasets
Ed25519
Cryptographic signing on every evaluation
286,510
ICU encounters analyzed (MIMIC-IV + eICU-CRD)
Engine running live Cross-validation complete: MI × elderly (eICU confirmation n=4,069, mortality 14.55%) -
The sepsis work

Pre-registered. Falsified. Published, including the null.

Falsification study · published null

A pre-registered falsification study of sepsis prediction benchmarks.

286,510 ICU encounters · MIMIC-IV + eICU-CRD (~209 US hospitals)
OSF pre-registered · medRxiv preprint · JAMIA Open in review

Labels were regenerated under three independent sepsis definitions; reported benchmark accuracy depended on which definition generated the labels. The pooled clinical claim did not survive the pre-registered protocol, so the null was published rather than shelved. That is the method: nullify our own result before anyone else has to.

Read the paper →
Replication · institution-type bias

Community-hospital sepsis workload: institution-type bias in ICU AI.

136,864 encounters · pre-registered replication
Academic-center-trained sepsis AI, measured in community hospitals

Sepsis AI trained on academic-medical-center data underperforms in community hospitals, and the gap is structural, not statistical. The institutions with the least governance capacity inherit the largest miscalibration.

Read the paper →

Sepsis prediction is also a locked AI Governor modality (SOFA + Sepsis-3 criteria), evaluated off-patient and signed. See all research →

Tools

Pick the trial that fits your work.

Tool 01

AI Governor

For CMIOs, AI governance committees, and clinical informatics teams.

  • Evaluate AI tool predictions against KDIGO, Padua, APACHE II, C-SSRS, and other published scoring instruments
  • Surface discordant verdicts with named failure modes
  • Cryptographic signature on every evaluation result
Trial includes pre-loaded demo cases for immediate evaluation. Bring de-identified model outputs from a tool you already use for an institutional engagement.
Request AI Governor trial →
Tool 02

Discovery

For research methodology, replication validation, and publication preparation.

  • Query an over-48-million-document research corpus with methodology-aware specialists
  • Surface contradictions between claimed findings and underlying evidence
  • Pre-registered methodology with hash-anchored attestation
Trial includes access to our shared research corpus. Bring your proprietary research questions for collaborative engagement.
Request Discovery trial →
Tool 03 · Pharma R&D · 30-day trial

RocSite Caliper

For pharma RWE and observational research teams.

  • Leakage detector scans every column before PSM runs, blocked inputs named by column, not processed into spurious results
  • Cross-references published literature and returns a signed finding with verified citations
  • Runs on your infrastructure, proprietary cohorts never leave your environment
Bring your own observational cohort. Public datasets available immediately on approval.
Request Caliper trial →

All three trials: 30 days, verified professionals only. Manual credential review within two business days.

Request trial access for an institutional engagement →
How it works

Three steps from credentials to results.

Submit credentials

Manual review within two business days for institutional verification.

Receive API key and dashboard access

Public datasets queryable immediately upon approval.

Run queries against pre-loaded data

When you're ready for engagement on your own data, contact us.

Validated across independent clinical resources
medRxiv OSF PhysioNet FDA 510(k) Pre-Submission MIMIC-IV eICU-CRD
Featured-finding deep-dive

Myocardial infarction, elderly, ICU.
Published estimate off by 168%.

Across MIMIC-IV and eICU-CRD, two independent datasets, observed ICU mortality for elderly MI patients is 21.5%. The commonly cited published estimate is 8.0%. That's not a rounding error. It's a clinical risk stratification failure that shapes bedside decisions and every model trained on the literature.

Read the full finding →
Condition × AgeMI × >70
Observed (MIMIC-IV)21.5% · n=2,161
Published estimate8.0%
Divergence+168.4%
eICU confirmationn=4,069 · 14.55%
Status● CONFIRMED
Evidence hash6bf704b3907a061b
After your trial

Engagement on your data.

Institutional engagements run YOUR proprietary data through our verified methodology. Engine License for proprietary corpora, Adversarial Validation for pre-FDA submissions, Clinical Partnership for collaborative research. Engagements are scoped to the institution. Contact us.

See services →