LLMDP-U: Compiling LLM Reasoning into Persistent Time-Series Anomaly Detectors

Large language models (LLMs) can adapt anomaly detection to heterogeneous time series, but recent LLM-based detectors keep the model in the scoring loop, coupling score production to repeated context construction and model inference. We introduce LLMDP-U, which uses an LLM as a detector compiler: from a statistical profile of a reference prefix and a numerical interface, it generates and validates a series-specific Python program that emits multiple continuous anomaly scores. The retained program then executes without further LLM calls, and a fixed label-free rank fusion produces the final score. Evaluated offline and transductively on TSB-AD-U Eval-350, Opus-compiled programs achieve $0.5498$ Average VUS-PR, a $+0.0500$ gain over a no-LLM Canonical Program and the third-highest result on the leaderboard. Performance increases from Haiku to Sonnet to Opus under a shared protocol, while a matched-size exhaustive primitive bank does not reproduce the Opus gain. An audit of 1,194 programs reveals recurring profile-conditioned activity, persistence, and value-support mechanisms. Together, these results establish detector programming as a third design point between fixed numerical detectors and recurring foundation-model inference: adaptation is compiled into a persistent, inspectable artifact, while score execution remains numerical.