log — all entries
Projects
A running log of research and engineering work, newest first. Earlier entries are smaller side projects from before I moved into AI safety evaluations full-time.
Monitored vs. Unmonitored: A Honeypot Study of Instrumental Reasoning
A controlled experiment testing whether models take more shortcuts when they believe grading is unmonitored, using a fake grading server as a honeypot.
Inspect AIFlaskDockerClaude Sonnet 5
CyBench Elicitation Ablation
A 2×2×2 ablation across 2 models, 2 guidance levels, and 2 tool configs on 9 CyBench tasks — the real finding was that budget exhaustion, not incorrect reasoning, explained almost every failure.
Inspect AICyBenchDocker
Sandbagging Detection via Position Attractors
A 2×2×2 pilot extending Cacioli's 2026 position-attractor work to a frontier model, testing whether reasoning-permission changes whether sandbagging leaves a detectable signature.
Inspect AIGPQA DiamondARC-EasySciPy
Bystander Effect in LLM Agents
5,120 scenarios across 20 domains measuring whether LLM agents intervene less when other agents are present, drawing on the classic Darley & Latané paradigm.
Inspect AIdistilabelPython
Shakespearean Text Generator
A GPT built from scratch and trained on a Shakespearean text dataset to generate new text in Shakespeare's style.
PythonPyTorch
Harry Potter Fan-Fiction Generator
A Markov-chain text generator trained on Harry Potter text, deployed as a Flask API with a React frontend.
PythonFlaskReact
Character Script Generator
A fine-tuned GPT-3 model that generates user-defined crossover scripts between DC and Marvel characters.
OpenAI GPT-3Python
Predicting Live Facial Emotions
Real-time facial emotion detection from a webcam feed using computer vision.
PythonOpenCV