# Model Experiment Boundary

> Training and model-combination ideas are research, not first-release capability.

- Site: TinyRustLM
- Canonical: https://TinyRustLM.mirust.com/model-breeding/
- Version: 1.8.2
- Updated UTC: 2026-08-30T01:36:19Z

## No evaluation leakage

Failed evaluation prompts, holdouts, expected answer concepts, and private outputs do not become training data. A rejected model may lead to a new source or contamination-controlled experiment with a new sealed holdout, never a lowered gate.

- Sealed holdouts
- No training on failures
- No evaluator rewriting to favor a model
- Honest reject is allowed