We build infrastructure for AI agents to code, deploy, and run research.

A language, compiler and runtime exploring how model calls, typed contracts and capability policy can live in the same program as ordinary computation. The Rust reference implementation includes a source-built CLI, an interpreter, an opt-in bytecode VM, editor tooling, bounded symbolic mathematics and local model inference. Developers can run retrieval-and-reasoning examples and inspect execution traces; the larger goal is to make an agent’s authority, uncertainty and evidence visible to the language rather than hide them in surrounding prompt code.

Pre-production: implemented slices, not a finished language. Production packaging, performance qualification, platform coverage and stronger isolation remain release blockers. Model judgments and passing tests are not formal proofs.

A research application written in Sema: choose a single search, a fact-extraction pass or a deeper investigation, then return cited facts, a summary or a structured report. It can also work from evidence supplied by the caller instead of searching again. A CLI and JSON API expose the workflow, with Markdown and optional PDF output. The deep mode tracks open questions and a statistical confidence estimate, stopping at an objective, confidence, time or iteration limit. That estimate controls the search loop; it does not certify the answer. Building the application also drove fixes in Sema’s generation, HTTP, configuration and citation machinery.

Experimental. Build the Sema compiler from source first; real generation and live web search require separate configuration. The opt-in offline demo uses a fixed HRV corpus, not evidence for arbitrary questions. PDF rendering also needs Python-backed rendering dependencies.

A PyTorch model that reuses one small transformer block across many iterations instead of giving every layer separate weights. Each pass rotates into a different latent subspace and receives the input again, testing whether more computation can substitute for a larger learned core. Researchers can vary loop depth, distil a teacher model and compare stability and task performance through Dentate. The momentum update has a reversible formulation, but that mathematical property alone is not evidence of constant-memory training or large-model-level capability.

A research prototype, not a production model. The roughly one-million-parameter figure counts the recurrent core, not the whole model. Broad-pool 100× knowledge compression remains unsupported; narrow-domain competence and reasoning depth are the hypotheses under test.

Training and evaluation tooling for small-model research. Runs retain resolved configuration, checkpoints and lineage, with commands to resume, fork and sweep experiments and a dashboard to inspect them. Alongside a real Apple Silicon fine-tuning path, the Spiral harness explores teacher distillation, narrow reasoning tasks and cross-window memory. The longer-term loop is Gauntlet scores → Dentate trains → Gauntlet re-scores; the fake-backed curriculum exercises that design without establishing that autonomous continual learning works.

Alpha, with two evidence levels: the broad Gauntlet curriculum is a fake-backed development pipeline, while MLX LoRA fine-tuning and the separate Spiral training/reasoning harness have real-model paths. Simulated curriculum gains are not measured improvements in a deployed model.