# Offline Model Breeding and Candidate Lineage

> Model breeding is an offline evidence pipeline: compatible parents, reproducible operators, receipts, manifests, runtime-smoke gates, eval sidecars, and selector admission.

- Site: TinyRustLM
- Canonical: https://TinyRustLM.mirust.com/model-breeding/
- Version: 1.3.0
- Updated UTC: 2026-06-29T22:59:10Z

## Not magic, not live mutation

The site presents model breeding as a governed offline artifact pipeline. Operators produce candidates, sidecars, or receipts. The browser only consumes staged and admitted artifacts after validation.

- Parent compatibility gate
- Lineage templates
- Operator receipts
- Candidate manifests

## Operator suite

The source describes blend, task-delta, sparse task-delta, pruning, DARE-style dropout/rescaled task-delta, federated/local-update, deterministic crossover, mutation, sign-aware merge, and seed-weighted parent-pool soup lanes. These are evidence-producing tools, not public quality claims by themselves.

- Blend
- Sparse deltas
- Prune
- DARE
- Crossover
- Soup

## Promotion path

A candidate must carry provenance and pass runtime-smoke; assistant-quality claims require evaluation sidecars with exact case evidence. Selector registries are the final browser route surface, not the start of the proof chain.

- Runtime-smoke
- Evaluation evidence
- Promotion ledger
- Selector admission