Autoresearch Loop

Use this method to improve one instruction atom without turning experimentation into an unbounded rewrite. The method owns selection, immutability, ratcheting, and review outcomes. A harness adapter supplies execution and isolation mechanics.

Select one eligible atom

Classify the proposed target before spending evaluation budget:

• L1 — deterministic behavior: route code, schemas, and mechanically provable behavior to ordinary deterministic tests instead of an autoresearch run.
• L2 — bounded instruction behavior: accept one mutable atom with a bounded diff, a frozen suite, and one accountable owner. This is the default eligible level.
• L3 — orchestration behavior: refuse multi-atom or interacting workflow changes unless a separately reviewed adapter contract provides an equivalent bounded oracle and isolation boundary. Identify a smaller L2 atom when possible.

Do not start when ownership, restoration, or the evaluation oracle is ambiguous.