Evolution boundary

Iteron includes types for a governed strategy-evolution control plane. It optimizes harness artifacts only and does not train or fine-tune a model or promote a live policy autonomously.

The model-task view of harness checkpoints explains why the selected policy bundle depends on both a measured model profile and a characterized task profile, and why transfer to another region needs new evidence.

Intended separation

The target architecture separates:

1. a fixed runtime trusted computing base;
2. replaceable strategy and world modules;
3. an isolated evolution control plane that produces immutable candidates.

Candidate producers may use search or other optimizer families. The retained supervised-fine-tuning, preference, GRPO, and RL method names are provenance labels for a harness artifact only. They do not authorize model training. The producer is replaceable; the promotion boundary is not.

ModelAdapter and ModelWeights are reserved wire values that manifest validation always rejects. Trajectory projection has no model-training target.

Promotion path

The planned evidence path is:

trajectory → governed dataset → candidate → held-out evaluation
           → shadow → canary → active

Every stage needs immutable identity, lineage, budget, consent, and rollback evidence. Evaluation fixtures and scoring policy must be reviewed independently from the strategy expected to improve the score.

Fixed human-controlled constraints

An evolved module may propose bounded actions. It may not:

• grant itself capabilities;
• relax security, durability, or budget policy;
• rewrite trajectories or evaluation evidence;
• choose its own promotion threshold;
• promote or roll back itself;
• expand data use beyond recorded consent.

Live self-evolution is intentionally outside the critical path until the runtime, effect, recovery, evaluation, and production-distribution gates are credible.