Publications — Reports & papers
Technical reports, in the open.
Our reports, experiment logs and code live in a public research repository. Papers move to arXiv as they are ready.
Oct 2026
Tree-Memory Architectures for Context-Free, O(log n) Language Models (v1 & v2)
Four candidate architectures, the open problems, and tested fixes: fresh radii, trained key geometry, conversion and from-scratch routes, consolidation and write trust.
Oct 2026
Chess testbed validation (exp07–exp09)
Rules, learning during play, recall of earlier moves and scaling to 1M addresses, with identical answers to full attention.
Oct 2026
Real-model conversion: root cause and routing keys (exp01–exp06)
Why bounds are loose on a pretrained model and what the first trained-routing-key attempt taught us.
Coming soon
Graph-Compressed Embeddings
Paper and package for anchor-plus-residual vector storage.
Coming soon
SHVQ-LR: semantic codebook compression
Method, experiment log and results for post-training weight compression.
Reports describe work in progress; numbers come from the testbeds stated in each report.