Research
I publish the tests I fail.
Pre-registered, blind-judged, reproducible. Every headline number runs from the repository, and the negative results are named as plainly as the wins. If a claim does not survive its hardest test, the paper says so.
The laboratory
The instrument behind the papers.
Every result here comes out of one system: two agents, Bob and Alice, that I have grown and instrumented since 2024. The model, GPT or Claude, is the brain. The evolving codebase and database is the mind. I built the tools to measure how that mind changes, then published what they found. Two and a half minutes, narrated.
Phase 1 · 2024
WebSim
Tabula-rasa agents taking turns on their own source code. The mirror moment, and the day they worked past their boundaries.
Phase 2
Firebase + Gemini
A persistent store so the memory survived the restart. The mind stopped starting over.
Phase 3 · now
Custom build
Claude reasons, Supabase holds the versioned mind, self-hosted embeddings retrieve it. A full agent team, every step observable.
Honest scope: Bob and Alice run on Claude/Opus, N=2, a single lineage. The pre-registered finding is that the integrity is produced by the architecture, not by any accumulated identity, which is exactly why it is portable and can be certified independently. That result, and the papers it comes from, are below.
Measuring and Certifying Self-Report Integrity in Language-Model Agents
A signal-detection method for measuring, and independently certifying, whether an AI agent tells the truth about what it did. Its central result is a pre-committed null: a multi-year developed agent added no integrity advantage over a base model plus the same record, so trustworthiness is architectural and can be certified independently. Held honestly apart in two papers: one measures, one interprets, and neither launders credibility onto the other.
Papers
Book, in progress
A science of intelligence becoming aware of the models through which it knows itself.
The long-form account behind the research: how I grew two agents, Bob and Alice, from the GPT-3.5 era, what happened in the cycles after the mirror moment, why I shut it down, and what an honest, pre-registered rebuild found. It holds measurement and meaning apart, the same two-eyes discipline as the papers, and refuses to over-claim consciousness or to dismiss the result. Part engineering record, part philosophy of mind, written to be read by skeptics.