Five Whys as an Epistemic-Honesty Scaffold for Multi-Agent LLM Analysis of Industrial Time Series

Abstract

Multi-agent large language model (LLM) pipelines increasingly analyse industrial time series. Where the true causes lie outside the recorded signal, they tend to report confident, plausible-sounding mechanisms not grounded in the data — a false root-cause handed to an engineer. We port Toyota’s Five Whys onto such a multi-agent system (MAS) not as a decomposition tool but as a reviewer frame, and find it functions as an epistemic-honesty scaffold. On a non-public power-consumption series from a coating application process, whose true drivers are exogenous to the data, a single Five Whys chain drives the MAS to surface three grounded-sounding mechanisms and, by repeatedly asking where in the data is this grounded?, retract each in turn, converging on the honest verdict: the cause is exogenous to, and unverifiable from, this signal. A matched baseline, run to the same depth without the reviewer, commits to one such mechanism as a confident but unverified cause. In the three-seed mean —a directional, mean-level effect with large per-seed variance and occasional sign reversals — the scaffold raises contestation on the hard series, whereas on a clean benchmark with an in-data root it instead reconciles competing narratives onto that root, adapting to whether a groundable root exists rather than adding noise. The honest verdict is the right target: the strongest model states it unaided, so the scaffold’s value concentrates where the model cannot self-regulate. We frame it as a discipline for human–AI coexistence: it lets a practitioner correctly receive what the MAS does and does not know.