A 311 lead notices the same complaint summary wording differs across identical retries. What LLM behavior explains this?

Identical 311 complaint retries that yield differently worded summaries reflect sampling non-determinism in LLM generation. Even with fixed prompts and inputs, next-token sampling can produce paraphrases that preserve meaning but change surface form—expected behavior, not evidence of prompt caching reusing outputs from similar 311 tickets. Linking summary differences to prompt caching reusing outputs from similar 311 tickets fails because caching serves stored outputs for matching inputs, which would make identical retries match rather than vary in wording. Interpreting wording changes as context compaction summarizing older complaint history fails because identical retries contain the same complaint with no older history to compact, and compaction condenses input context rather than rewording the model's output across retries. Treating retry variation as scheduled model version migration changing civic summary phrasing fails because identical retries hit the same deployed model version; migration occurs across deployments over time, not between repeated requests. Exam caveat: non-determinism complicates regression testing and resident-facing consistency; pin temperature/seed where the product requires stable phrasing, or compare semantic equivalence rather than exact strings. Operational check: replay the same complaint payload N times, log summary digests, and document whether variance stays within an accepted paraphrase band for 311 quality review.