Conformal prediction under cross-period exchangeability breakdown: a causal-admissibility M&V cascade for LNG receiving terminals

Liquefied natural gas (LNG) receiving terminals require measurement and verification baselines that support energy-performance assessment under the IPMVP and ISO 50006 frameworks. Post-hoc interpretability methods explain a fitted model but do not themselves document an independent pre-fit input review. This paper develops a six-stage cascade with a separate pre-fit review. A pre-prediction Causal Admissibility Screening audit decides admissibility from the raw data alone and emits a log that an external verifier regenerates without any trained-model checkpoint; causal discovery then restricts the forecaster's pooling readout to the recovered ancestors, a Gaussian-process residual model with weighted-block conformal calibration supplies the prediction intervals, and a topology-weighted scorer ranks process nodes for non-routine-event hypotheses. On a 180-day LNG simulation the cascade attains strong point-prediction accuracy on specified reporting subsets and 3/3 conditional top-3 candidate coverage under an explicit limited-power caveat. Its central finding is negative: under the short calibration window, conformal coverage collapses across periods, a cross-period exchangeability breakdown traced to a nonconformity-score distribution shift. The tested naive adaptive-conformal repairs restore coverage only by suppressing acute faults, whereas a recency-weighted non-exchangeable variant escapes that trade-off under chronological online replay. A paired post-fit residual calibration then brings the reporting row within the study's point-accuracy targets without reopening the admissibility decision. A supplementary test on 90 independent fault/healthy pairs reveals residual false alarms and rankings identical to a constant reference. Three recommendations follow: a baseline covering relevant operating and seasonal conditions, an independent admissibility audit log reproducible without the trained model, and direct-state SCADA templates at each major processing node.