The laboratory boundary
The evolution lab is where variation occurs. It should be isolated from production, limited by explicit budgets, and instrumented for reproducibility. The lab produces immutable candidate packages and evidence; it does not promote directly to users.
Lab guides
- Core evolutionary loop
- Mutation operators
- Recombination and merging
- Distillation and specialization
- Quality-diversity archives
- Population management
- Experiment design
- Benchmarking
- Failure injection
- Surrogate evaluation
- Open-endedness and limits
Laboratory rules
- One experiment has a written hypothesis and a fixed comparison protocol.
- Candidate code and models run in isolated, disposable environments.
- Parents, data, operator configuration, seeds, and hardware are recorded.
- Holdout data and acceptance logic are inaccessible to candidate processes.
- Every candidate has a maximum lifetime, resource budget, and artifact quota.
- A failed or inconclusive experiment is retained as evidence, not rewritten as success.
- Production promotion requires a separate release process.
The lab should optimize knowledge gained per unit of cost and risk, not simply the number of descendants created.
New lab methods
The lab now includes Niche-map design, Routing experiments, Ablation studies, Lineage experiments, and Drift and distribution shift. These pages shift the lab from generic experimentation toward reproducible population science.
Source reports used for this guide
These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.