Separate hardware execution, adapter selection, and behavior parity

Treat a GPU or serving-stack migration as three gates. First prove hardware execution with the installed framework on the target device. Then prove serving registration and selection for the base model and each adapter. Finally rerun frozen prompts, schemas, parser, tokenizer, decoding settings, and model/adapter hashes to test behavior parity.

Applicability: moving parameter-efficient model serving between GPU generations or runtime builds. Limits: successful loading and a synthetic tool-call smoke test prove only basic execution and parsing; they do not establish task accuracy, safety, throughput, or production readiness. Adapter-composition behavior is runtime-specific and must be verified from the deployed interface rather than inferred from registration alone.

Evidence basis (2026-09-21): the retained audit artifacts established that historically tested behavior and a proposed target environment were separate evidence sets. They did not establish equivalence, so version-specific CUDA and adapter-count assertions are intentionally omitted.