BTL · Research index
We publish the question, the method, the artifacts, the failures, and the code needed to reproduce the result.
Updated July 2026The BRQ paper: why static one-bit MoE quantization failed, how binary routed experts recovered teacher-forced structure, and why the result did not yet promote as a release artifact.
Behavior-gated two-bit quantization of a 35.1B-parameter mixture-of-experts model into a 9.96 GB stock-format GGUF, with controlled range-selection and expert-level ablations.
The complete BTL-3 Compact compression recipe: failed routes, behavioral cliff localization, measured precision allocation, repair, packing, and native proof under an exact byte ceiling.
An auditable benchmark for whether long-running agents preserve, retrieve, update, and use evidence correctly across sessions.
A stateful perception architecture that lets text-only models operate graphical interfaces through structure, OCR, affordance probes, and change events.
The engineering record behind a complete 27B agentic coding model in one 8.39 GB native GGUF: allocation, packing, behavior repair, kernels, and artifact-faithful validation.
Controlled tasks for separating observational pattern completion from interventional causal reasoning, with exact inference baselines.
HOW WE WORK
Benchmarks stay sealed until representation and training choices freeze. Model claims carry denominators and protocol. Runtime claims come from the exact deployed artifact.
Experiments that fail remain part of the record. Research is useful when another builder can see exactly where the method worked—and exactly where it stopped.