Engineering & research benches
Engineering workbenches and provenance docs
Offline research benches for open-weight models, with a documented decision for every step.
Published · Updated
- Open-weight models
- Local inference
- Markdown docs
- TypeScript
What it is
studio.gkheart.ca hosts research benches: harnesses I built for running, measuring and training open-weight models offline. Each experiment is a structured workflow, and each decision is written down with its reason.
High-context, low-reasoning pipelines
Smaller local models struggle with open-ended reasoning but do well when given precise context and a narrow task. The benches test exactly that: they front-load documentation, schemas and examples, split work into small steps, and measure how reliably a modest model completes each one.
Provenance
The goal is that a project can be recreated from its own documentation: readmes, design docs, schemas and recorded decisions. If a feature cannot be explained and rebuilt from those records, the work is not finished.
- Every build step has a recorded decision and reason
- Experiment configs are saved next to their results
- Human validation before anything is merged