Research

Pre-registered. Peer-reviewable. Published with null-result transparency.

All RocSite research follows the same discipline: pre-register the hypothesis, lock the analysis plan, publish the result regardless of direction.

Primary medical AI research.

Three pre-registered medical AI studies sit under one OSF umbrella project (osf.io/3ws8g). Each hypothesis and analysis plan was locked before any data access. Results are published regardless of direction.

  • Pre-registered falsification testing of sepsis prediction models 286,510 ICU patients · MIMIC-IV + eICU-CRD · methodology reviewed by an established sepsis informatics researcher · primary cohort returned a null result, published anyway Read on RocSite →  ·  medRxiv preprint  ·  OSF pre-registration  ·  GitHub
  • ICU mortality miscalibration in older adults 201,905 ICU encounters · pre-registered 2026-03-10 · observed-vs-published divergence in elderly mortality: +66.3% (AF), +131.7% (diabetes), +168.4% (acute MI) · submission pending OSF umbrella →
  • Community-hospital workload & cross-institution AI performance 136,864 encounters · pre-registered 2026-02-12 · cross-institution AUC loss of 11–19 points when models trained on academic cohorts are applied to community-hospital data · submission pending OSF umbrella →

Two of the three studies are pre-registered but not yet submitted. They will appear with their preprint and code links here as each clears submission. The umbrella project (osf.io/3ws8g) carries the locked timestamps in the interim.

Methodological proofs of concept.

The same adversarial falsification methodology used in the sepsis study has been applied to two undeciphered historical corpora as methodological proofs of concept. These are presented as falsifiable statistical findings, not solutions. The findings live at their dedicated research archives, where the full methodology, falsification conditions, and corpus-level data are documented.

Why we frame these as findings, not solutions. Both corpora have a long history of confident-sounding claims that did not survive scrutiny. We are not adding to that pile. The methodology is what generalizes; whether any particular interpretation of the underlying corpus is correct is a separate question, decided by other people doing other work. The dedicated sites document exactly what would falsify each finding.
Get in touch

Critiques and replications welcome.

If you have read the work and want to point out something we got wrong, please email Adam directly. Pre-registrations are public, code is public, and replication is welcome.