RENDERING…

Hedronite

Hedronite Foundry

Where We Forge the Tools We Use.

Hedronite Foundry is where we build the runtime that carries our operators and the specialist models they call on. Distillation pipelines convert Hedronite-canonical corpora (Atrium knowledge and Akasha session traces) into training data for Gemma 4 variants. Continuation training and supervised fine-tuning produce specialist models matched to the work each operator does. claurst, a clean-room runtime written in Rust, holds those models in production. We do not pursue artificial general intelligence; we build specialists matched to the work, and the runtime that runs them.

Distill

The first stage. Two flows. Atrium carries the canonical knowledge a Hedronite specialist needs to recall; Akasha carries reasoning traces from real operator sessions. Distillation converts both into training data without flattening the distinctions between them.

  • Atrium knowledge distillation: structured compression with retrieval-shape preservation
  • Akasha reasoning distillation: chain-of-thought capture across operator sessions
  • Corpus versioning, provenance tracking, and DataOps quality stamps before Forge

Forge

The training stage. Continuation pretraining on Gemma 4 variants against the distilled corpora, followed by supervised fine-tuning on curated chain-of-thought sequences. Where a frontier-scale lab spends ten thousand GPUs, we spend ten, on material the frontier-scale lab does not have.

  • QLoRA fine-tuning on Gemma 4 base variants
  • Continuation pretraining and supervised fine-tuning passes
  • Distributed training across the Hedronite hardware mesh, checkpoint discipline maintained throughout

Tune

The refinement stage. Where the forged model is shaped into the specialist the operator will actually use. Evaluation against Hedronite-canonical benchmarks, A/B testing, and the final shape-checks before the specialist is mounted into production.

  • Hedronite-canonical benchmark suites for domain accuracy, reasoning fidelity, refusal correctness
  • Evaluation under the QA + DataOps lane; adversarial validation under the red-team lane
  • A specialist does not ship until it beats the model it replaces on every named benchmark

Mount

The runtime stage. claurst, our clean-room Rust runtime, holds the model in production with a Kosha buff, a three-layer memory architecture, and a tool-call classifier. The Mini-3 operator cohort runs on claurst today; specialist Gemma 4 deployment lands on the same runtime once Tune ships.

  • claurst: clean-room Claude Code rewrite in Rust, shaped to the work our operators do
  • Kosha buff, Three-Layer Memory architecture, YOLO tool classifier
  • Production instrumentation and drift detection for every model under management

We do not sell our specialists, license our runtime, or take outside engagements for fine-tuning. Hedronite Foundry is where the rest of the operation gets its tools.