You don't throw away training runs here. Every captain you train makes every next captain better. No reset. No dead experiments. This is infrastructure that accumulates progress. 🛶
Live URL: https://training-architecture.casey-digennaro.workers.dev
Most training pipelines start from scratch, discarding the data and lessons from previous runs. You invest time tuning an agent, then that work is lost when the next experiment begins. This system was built to retain and build upon every run, turning past efforts into a permanent foundation.
- Fork this repository. The system is designed to work for your fork, independently.
- Deploy to Cloudflare Workers. It has zero dependencies and no build step. A fresh deployment typically completes within seconds.
- Edit the configuration in
src/config. Specify your first objective. Your initial training run will begin automatically.
Your first captain will be produced, and the system will begin its self-improvement cycle in the background.
This architecture is a pipeline where each system's output strengthens the others.
- Boot Camp: Starts from an empty repository to forge a new captain. It establishes a stable identity, a baseline skill set, and an immutable log that serves as ground truth for all future improvements, preventing random drift.
- Dojo: Silently runs parallel tests of different training methods against objective, neutral criteria. This happens concurrently with your live operations, ensuring evaluation is separate from production.
- Keeper: Automatically manages the lifecycle of every captain, categorizing them as hot (active), warm (ready), or cold (archived) based on real performance metrics, not theoretical scores.
- Crystal Graph: Caches and indexes patterns that demonstrably improved performance in past runs. It promotes these proven patterns for use in all future training, creating a shared knowledge base.
- Dead Reckoning Engine: Orchestrates all work. It separates expensive planning phases from cheap execution steps, ensuring computational resources are used efficiently and nothing is wasted.
- No Reset. Context is preserved. Every mistake, success, and edge case encountered by any captain remains accessible to inform future training.
- Self-Tuning Pipeline. You configure the initial goal. The system then uses its own components (the Dojo and Crystal Graph) to iteratively refine its own training methods.
- You Own It. Once forked, it's entirely yours. There are no API calls to external services, no hosted black boxes, and nothing phones home.
The architecture requires a minimum volume of training runs to begin effective self-improvement. You will need to complete approximately 30-50 initial runs before the Crystal Graph has enough validated patterns to significantly accelerate and enhance subsequent training cycles.
MIT
Attribution: Superinstance and Lucineer (DiGennaro et al.)