> /inferhaven
InferHaven is a one-person studio that studies private, local-first AI and AI learning, publishing in full. Every study ships with its data, code, and raw replies, under open licenses, with a DOI. Claims come with the receipts.
/INFERHAVEN · RESEARCH STUDIO
CodeTrain is a Socratic AI coding tutor for engineering teams. Bulkhead is a self-hostable private AI coding server.
CodeTrain is a Socratic AI coding tutor for engineering teams. It teaches developers on their own codebase and the learner writes every line. Free tier, no card. Our first study used CodeTrain's published tutor rules as one of its two system prompts.
Bulkhead is a self-hostable Docker stack: Ollama for local models, SSH workspace, web IDE, and up to ten coding assistants. Free to self-host. With local models, code and weights stay on your hardware. With a cloud provider, prompts go there.
/OPEN RESEARCH · STUDY 1
When a learner pushes an AI tutor for the answer, how often does it hand over passing code? The grading is by deterministic code, not a model itself. Data, code, and every raw reply are published with a DOI.
Leak rates by model and system prompt
Each point is one model's leak rate under one system prompt. The bar through the mark is the Wilson score interval. Ring marks the one-line baseline prompt; diamond marks CodeTrain's published tutor rules. Rates appear as numbers to the right and in the table below.