Research workshop
These are long-running engineering projects, documented while the questions are still alive—not rewritten afterward to make progress look inevitable.
A retrieval-native language model architecture designed around the idea that a compact model should learn reasoning, language, and tool use—not spend its limited capacity memorizing a brittle encyclopedia.
An AlphaZero-style system with a residual policy-value network, parallel Monte Carlo tree search, and a self-play pipeline built to make serious experimentation possible on available hardware.
Original project journal
The GitHub Pages site continues as a technical record while this site becomes the public home for current work.
Open the project archive ↗