Answer first
A real ModelBreeder lab should expose small, composable commands that write file-backed evidence. The commands should use UTC timestamps, deterministic seeds, artifact digests, and positive fitness language.
Suggested commands
| Command | Purpose |
|---|---|
slm_manifest | Read or write model package metadata. |
slm_validate | Verify tensor layout, tokenizer identity, adapter compatibility, and checksums. |
slm_fitness | Run evaluation cases and write a FitnessVector report. |
slm_novelty | Compare behavior descriptors against the archive. |
slm_merge | Plan or execute compatible multi-parent merge recipes. |
slm_release | Build a release packet with rollback target and next action. |
Example
COMMAND slm_fitness \
--model ./registry/models/legal-summary-v4.slm \
--cases ./eval/legal-summary.cases \
--fitness-report ./registry/evidence/legal-summary-v4.fitness.json \
--created-at-utc 2026-06-28T00:00:00ZNaming rule
Use terms like fitness-report, evaluation-scope, next-action, candidate, champion, and specialist. Keep warning-heavy terms out of the main lab interface. The extended negative case belongs on Cognivirus.com.
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.
Implementation directivesControlled Evolution of AI Models: Site Copy, Schemas, and Breeding LoopImplementation directive · 25.7 KB
Browser and edge runtimeArchitectural Advancements for Zero-Dependency In-Browser Large Language ModelsEmerging implementation architecture · 54.0 KB
Implementation directivesModelBreeder Site and Theory: Controlled Evolution and Positive ImplementationImplementation directive · 25.7 KB