Research

Evaluation science for high-stakes AI.

PSAIL's research asks a single question in four different ways: can this AI system be trusted where the cost of error is a wrongful outcome?

Method programs
Domains

Methods across high-stakes domains.

The four programs are method axes, not domains. PSAIL proves them first in public safety, then extends them to adjacent settings where the cost of error is just as high — including national and economic security.

Public SafetycoreNational & Economic SecurityactiveHealthcarefutureRoboticsfuture
The Research Map

Methods, projects, impact.

One page, from foundations to impact: data and legal constraints feed four method programs, each program drives a project, and each project produces its own papers, protected methods, and real-world impact — across public-safety and security domains.

PSAIL research system map Four method programs — Evaluation, Governance, Assurance, Systems — each drive projects that produce papers, methods, and impact. The methods are proven in public safety and extended to adjacent high-stakes domains including national and economic security. FOUNDATIONS METHOD PROGRAMS PROJECTS OUTPUTS & IMPACT baselines constraints admissibility personas benchmarks controls tech-leakage assured data simulation edge AI deploy Human-expert baselines Law & proceduraljustice 1M syntheticpopulation AIEvaluation AIGovernance AIAssurance AISystems PoliceBenchreliability benchmark Public Safety AIgovernance framework Economic Security AINat.-Security domain · early GABMagent-based simulation EV-Hub edge AIAccepted · edge-AI networks Police for Everyonecivil-petition (민원) AI · citizens LLM-as-judge paper · In review expert-graded reliability suite Trilemma paper · In review policy & institutional directives Fraud & disinfo GABM · In review Persona validation · In review Protected methods · in filing National deploy · ₩10.67B · in dev. civil-petition platform for citizens DOMAINS Proven in public safety, extended to adjacent high-stakes domains. Public Safety National & Economic Security Healthcare Robotics future
Open science & disclosure

How we work.

Publish on acceptance

Papers appear here only once accepted. Work under review or in preparation is marked in progress — never dressed up as published.

Disclose after filing

Inventions stay confidential until filed. The Research Map shows the method stage, but never unfiled claims or content.

Validate against ground truth

Synthetic populations are calibrated to national statistics before use, and evaluation is anchored to human-expert baselines.