Why Predict What You Can Preserve? Controlling Generative Scope in LLM State Updates

LLM agents increasingly modify persistent code and structured state, yet even a small requested change is often delivered as an entire rewritten successor. This makes the model regenerate source content that is already known, exposing untouched fields, delimiters, and code to omissions or incidental changes. Output validation can ensure a well-formed candidate without ensuring fidelity of unchanged state; compact edit formats restrict model-authored content but can fail at address resolution or value encoding. We therefore treat preservation as a relation between an authoritative source and its successor, rather than as a property of an output in isolation. We define generative scope as the source region whose successor content the model must provide, and generative exposure as the unchanged portion of that region. Our method binds a supplied edit to the source, allocates model versus runtime responsibility for the successor, and composes accepted local replacements with unchanged source content. Under sound source and address binding, this composition preserves the untouched frame by construction, while leaving replacement correctness and applicability as separate obligations. This formulation makes preservation and edit-format failures measurable without asserting a model success law. With the correct replacement and address supplied, local Patch improves byte-exact fidelity over full Rewrite by 40.0 and 39.1 percentage points on QuixBugs (40 tasks) and all eligible HumanEvalFix-Python edits (133 tasks); released-test gains are only 2.5 and 0.8 points.