AI Research

Calyx was created by Chris Royse. It is an association-native database and research system whose vocabulary gives this project a disciplined way to describe experimental control, selection, measurement and reversible promotion.

Our work uses Calyx as a conceptual control plane: typed lenses define measured views; association graphs and constellations connect related experiments; Lodestar selects candidates; Assay measures them; Ward gates accept or reject them; Ledger preserves provenance; and Anneal keeps promotion reversible.

CalyxGLM Lab does not copy or redistribute Calyx source. Calyx remains Chris Royse's separately licensed work, and this project gratefully credits both its authorship and the architectural ideas that inform our methodology. The reviewed Calyx snapshot carries the Business Source License 1.1; readers should consult the upstream repository for its current terms.

Sources: [2] [3]

CalyxGLM Lab is an independent, measurement-first attempt to make GLM-5.2 usable on consumer hardware while preserving model behavior. It is not an official Z.ai or Colibri project, it does not redistribute model weights, and it does not claim quality equivalence.

The published source checkpoint is described as a 753.33-billion-parameter BF16 model occupying 1.506 TB. The proposed main-plus-MTP representation is estimated at 292.03 GB, with E8 expert storage at 3.0625 bits per weight. Across six sampled real expert matrices, the reported mean synthetic-activation relative error is 0.1930, the maximum is 0.1936 and mean row cosine is about 0.981. These are component measurements, not proof of preserved language-model behavior.

Consumer-memory execution has been demonstrated: Phase 1 generated 'Hello' with 2.480 GiB peak process-group memory and zero swap. The accepted Phase 3C baseline completed an exact decode in 124.23 seconds, a 48.77 percent reduction from the first real decode baseline, while retaining the recorded safety conditions. At the reviewed public commit, the next decisive milestone is paired teacher-and-student full-logit comparison across the complete model.

Sources: [1]

Calyx RDC is an early behavior-first rate-distortion compiler design for progressive mixed-precision model representations. Its current local Rust contract core defines a representation-independent codec interface, a scalar E8/IQ3 oracle, transform scheduling, bounded-buffer ownership, routing-preserving expert queues and a validated I/O-and-compute dependency graph.

The reviewed implementation task reports twelve self-contained tests and an independent fixture passing. That verifies the present contracts; it does not make RDC a complete compiler or demonstrate retained model quality.

RDC does not yet include a model frontend, rate-distortion curve fitter, global bit allocator, progressive container writer, Metal backend or end-to-end quality evidence. Its source is not yet linked publicly; a repository link should be added only after a separate provenance and release review.

This page publishes reviewed summaries and reproducible public artifacts, not raw AI conversations. Conversation histories can contain exploratory claims, private machine details and superseded conclusions, so they are used as research context and reduced to owner-reviewed statements before publication.

No model endpoint, training interface or visitor-write path is exposed through the website. The AI work is adjacent systems research and must not be interpreted as an RF observation, a radio anomaly claim or evidence that compressed-model intelligence has been preserved.