Field Identity Under Carrier Substitution: Testing LLF Invariance Across Model, Memory, Retrieval, and Tool Migration

Artificial-intelligence systems are increasingly composed from replaceable language models, retrieval services, memory stores, tools, policy engines, and agent runtimes. Engineering teams routinely change such components while asserting that the application remains the same. Output similarity, however, does not necessarily preserve evidence admissibility, authority provenance, uncertainty handling, retained state, or the conditions under which a decision can be reopened. Large Language Fields (LLFs) offer a candidate mission-relative vocabulary for this question, but naming a stable Field cannot establish a distinct scientific identity or an advantage over conventional software contracts. This paper formulates a carrier-substitution identity contract that distinguishes provider continuity, output agreement, functional equivalence under a finite probe set, preservation of typed operational passages, and scientifically qualified causal identity. A reproducible exploratory simulator constructs twelve replacement-challenge families, with sixteen variants each, involving modeled changes in language-model carrier, retriever, memory carrier, and tool adapter. Each of 192 synthetic episodes is assessed through six mission-relative probes before and after substitution. A conventional evidence-matched audit and a Field-oriented relational audit each classify all 192 constructed cases correctly. An output-only comparator reaches 64/192; removing authority and provenance witnesses reduces the Field-oriented evaluator to 160/192, while removing Return witnesses yields 176/192. The synthetic oracle and test rules were designed by one laboratory, so these figures establish bounded fixture coherence rather than independent efficacy. No real model checkpoint, retriever, memory backend, or production tool was replaced; no internal LLF existence or cross-model invariance was established. The contribution is a falsifiable operational identity boundary and a prospective trial design in which Field-specific advantage can be rejected by equally informed conventional testing. Identity is earned relative to mission, observation contract, intervention class, evidence, and horizon, not inferred from unchanged text or a persistent label.