The four service tiers above are domain-agnostic. The published track record is in clinical AI;
the methodology, pre-registration, cross-dataset replication, adversarial debate, cryptographic
attestation, applies anywhere a quantitative claim has high consequences. Engagements outside
clinical AI are scoped on request.
Primary credential
Clinical AI
Audits of clinical prediction models, ICU mortality estimates, sepsis benchmarks,
institution-type bias, and outcome miscalibration. Three published studies on MIMIC-IV
and eICU-CRD; FDA-track work in progress. Cohort definitions and analysis code are
peer-reviewable; confirmed findings ship with reproducible bundles.
To engage, pick a tier above based on scope: Engine License for ongoing surveillance,
Adversarial Validation for pre-FDA falsification testing, Clinical Discovery Partnership
for joint cohort work, or Evidence Integrity Score for a single-paper audit. Contact
for scope conversation.
By engagement
Financial Models & Fraud Detection
Audits of backtests, risk models, fraud-detection classifiers, and signal-generation
pipelines. The same falsification framework surfaces lookahead bias, label drift,
survivorship, trade-execution leakage, and out-of-distribution failure modes that
collapse model performance once a strategy is live or once an adversary adapts.
To engage, the work begins with a scope conversation, then a pre-registered protocol
(hypothesis, data definition, falsification criteria, planned analyses), an
independent audit on held-out or client data, and a citable report with cryptographic
attestation. Available by engagement. Contact for scope conversation.
By engagement
Legal & Regulatory AI
Audits of AI used in high-stakes legal and compliance decisions: case-outcome models,
contract-review classifiers, regulatory-filing assistants, sanctions screening, and
reasoning-chain validation. The Adversary surfaces case-volume and venue-selection
bias, failure modes that disappear under standard accuracy reporting but matter once
the model is in front of a regulator or a court.
To engage, the work follows the same pattern: pre-registered protocol, independent
audit (on hash-anchored extracts when the corpus is privileged), and a citable report.
Available by engagement. Contact for scope conversation.
By engagement
Applied Research Validation
Pre-registration, replication, and methodology audit for any quantitative claim where
the cost of being wrong is high, published-paper integrity scoring, pre-submission
review for journals or institutions, replication of headline findings on independent
data, and methodology audits of forthcoming work before public release.
To engage, scope a conversation, then a pre-registered protocol with falsification
criteria, an independent run on the relevant data (or on equivalent data when the
original is private), and a citable report. Available by engagement. Contact for
scope conversation.