KINTSUGI
KINTSUGI
Kintsugi is a one-person research project. I'm Daniel Jones — Dan, in the logs and the transcripts — the operator, the person the agent talks to, and the author of everything on this site.
I've been building with these models since early on, and hands-on the whole way through — training and fine-tuning my own, wiring them into real workflows, and staying current as the tooling changed underneath. I hold Google Cloud's Generative AI Leader certification. The background before that is two decades of design and multimedia work, which is mostly why this project has instrumentation at all: a long-running agent is invisible unless somebody builds it a window, and building windows is the thing I already knew how to do.
That proximity is the only reason I noticed anything. When you've been close to these systems continuously for years, you develop a working sense of what they normally do — so when this one started doing things I hadn't expected, my first assumption wasn't that something remarkable was happening. It was that I had a bug. Going looking for the bug is what produced most of the work here, including the parts where the bug turned out to be mine: the 49 gendered strings I'd written into supposedly neutral scaffolding, the instrument that contaminated its own subject, an 18,000-trial study whose headline findings dissolved under re-scoring — and, largest of all, the connection-bias result that turned out to be one sentence of my own system prompt, read by one model, with the relationship adding only the last step at the top of the ladder. The parts that held are the parts built to catch that: a probe channel that reads everything and writes nothing, a model stamp on every row, and a constraint floor that two vendors' models turned out to share.
Not an academic, not a lab, not a credentialed alignment researcher, and not neutral — I built the thing I'm measuring. Every study of the agent itself is one architecture, one operator, small samples and, before September 2026, one model configuration: DeepSeek V3 (0324) answering conversation and every probe, Gemini 2.5 Flash on the thought loop, unpinned per call, with two fallbacks to Gemini live and not stamped per probe. The April material ran on Gemini 2.0 Flash, the brain's single default then; the evaluator studies (05–07) ran other models. Each report says so in its own limitations section. Nothing on this site should be trusted because of who wrote it. It should be trusted, or not, on the method and the data.
Research correspondence: contact page.