Projects
-
Predicting false financialcollusion alerts under incomplete information
Simulated a market where a standard screen's false-accusation rate for collusion rises 5.8% → 100% as auditor information degrades. Built a simple analytical model that predicts these rates closely, without being fitted to them.
-
Detecting agentic collusion in public logs with statistical methods
Used statistical methods to detect an LLM-agent swarm from public edit logs alone. The weekly check flags the three largest incident weeks and none of the 2,236 earlier wiki-weeks (2016–2025).
-
A Calibrated Two-Level Detector for Secret Loyalties – Application to an LLM
Designed an evaluation framework, then measured its false-positive rate on a clean model. Using matched-control calibration, cut false positives from 12/16 to 2/16.