Plausible Patients, Impossible Populations
An epidemiological audit of LLM mental health simulation: 28,800 generations across 120 demographic cohorts, scored against survey-weighted NHANES anchors.
I am an independent AI evaluation researcher in Sydney. My evaluation pipelines have run
more than 150,000 model generations and judge calls, and the finding repeats across nine
studies: large language models that pass binary safety checks still flatter
users, flatten the populations they simulate, and misreport the variables that drove
their own outputs. Two sit on arXiv as preprints, and every study ships with a public,
receipt-backed repository. I also write fiction and essays, none of it with AI.
Featured Research
The Granularity Gap
A multi-dimensional cross-generational audit of sycophancy in Gemini models, showing that an LLM judge's binary verdict captures only 29% of the variance in its own severity scores.
Telling More Than They Can Know
A four-model audit of whether LLM self-explanations disclose the demographic drivers of their psychiatric-instrument scoring, showing that models name demographic features but credit them with a fraction of the score variance they actually drive.
Featured Writing
The Basilisk
An experiential psychological thriller set in the shadow of Roko's Basilisk. Menote wakes to find everyone dead and his memories gone, drawn toward a future he cannot escape as his mind fractures into Greek Titans and the prose unravels from modern literature into Homeric bicameral writing.
The Second Tree
A philosophical travel journal spanning two and a half months across Prague, Kutna Hora, and Florence, capturing a fluid pursuit of gnosis while pressing against Christianity, Nietzsche, and the noise of modernity through the specific resistance of place.
Featured Projects
Ingenuity Readiness Evaluator
A commissioned agentic RAG pipeline that scores major organizations against a leadership framework, producing reports of roughly 110 individually cited findings per company with verbatim-quote discipline at every score.
Python • Flask • React • Claude API • Agentic RAG • Headless Chrome PDF
ClaudeMind
A self-maintaining personal memory system: vector and graph indexes over an Obsidian vault, kept current by autonomous Claude Code loops, gated by audience-scoped privacy profiles, and measured by its own evaluation harnesses. Portable release on GitHub.
TypeScript • FastEmbed (ONNX) • SQLite + sqlite-vec • MCP • Claude Code • Obsidian
Mimesis MCP
An offline, local-first MCP server that drafts in a target author's voice, and an evaluation loop built to find out whether it actually works. Hybrid retrieval and stylometric profiling, plus rhetorical-shape detectors built after blind testing showed the surface checks were not enough.
Python • SQLite FTS5 • FastEmbed (ONNX) • FastMCP • PowerShell • FastAPI