Action-Conditioned Latent Dynamics for Mineral Processing: A World-Model Engine, and a Measured Account of Where It Earns Its Cost

Abstract

A forecaster maps history to a future. A world model maps history and a proposed action sequence to a distribution over futures. That single difference is what makes counterfactual questions, planning and regime detection expressible at all, and it is the difference this paper is about. We describe latentplant, an open (MIT, PyPI) engine for action-conditioned latent dynamics on industrial process time series, and we report what it actually achieves on mineral-processing data under an evaluation protocol designed to make failure visible. The engine contributes an ingestion contract that declares each tag's role and the cadence at which it truly updates, a leakage-safe windowing layer, a ladder of rungs from persistence to an ensembled recurrent state-space model, open-loop multi-horizon scoring under both point and proper probabilistic rules, split-conformal interval correction fitted on a held-out calibration split, sampling-based planning, and two partially observed reference environments with exact ground truth. The results point in three directions and we report all three. On a one-dimensional thickener PDE whose entire concentration field is hidden behind four instruments, the recurrent latents lead on the instrument that reflects that field: RMSE 0.3798 for the single latent and 0.4118 for the ensembled one on the bed level, against 0.4947 for the best non-recurrent rung and 0.5368 for persistence. Pooled scale-free over all four channels the top of that table is nearly tied instead (0.5108 for the probabilistic ensemble against 0.5115 for the latent), because the other instruments are easy for every rung: the same errors pooled in raw units would have reported a sweeping win that belongs to one channel. The two recurrent rungs then separate on confidence rather than accuracy, which is the case for building the ensembled one: alone, the latent covers 0.556 of outcomes at a nominal 90 percent (0.336 on the bed level) and ranks 5 of 9 by a proper scoring rule; ensembled over four initializations it covers 0.821 (0.991) and ranks 6, so what the spread buys is a usable interval rather than a better rank. On a grinding circuit that is smooth and near-linear about its operating point, the best rung by CRPS is DeepAR at 0.0768, with a standardized ridge VARX 3 of 9 at 0.0928; wherever a linear model places ahead of a latent world model, that is a result this evaluation exists to be able to report. On a real iron-ore flotation record, scored on the laboratory assays the case declares as its targets, the ladder does beat holding the last value at hourly cadence and the rung that does it is linear, while at minute cadence no rung beats it on point accuracy. The minute-cadence case had stated before the runs that such a failure would reject the "hourly averaging destroyed the signal" explanation, and it did, so we measured the mechanism: the principal exogenous disturbance is frozen for roughly nine hours at a time and does not move within a sixty-minute horizon at all. Scored over every channel rather than the declared assays, the same rollouts read the other way, because the fast process tags are nearly persistent; the scope of a score is part of its claim. On Tennessee Eastman, the one richly excited record in the matrix, every learned rung except the linear one clears persistence, and the best rung by a proper scoring rule is DeepAR at both horizons, at CRPS 0.3593 and 0.3767, covering 0.883 and 0.875 of outcomes against a nominal 0.90: on the simulated dynamics the best-scoring rung is also the one whose interval can be used. Its surprise stream detects 13 of 20 fault classes at a measured 1.02 percent per-sample false alarm rate while all three known-undetectable leakage probes correctly go undetected. We argue that the third result is the most useful one. It is a limit of that record at that cadence rather than of the method, and it is why the reference environments ship with the engine: the gap between what a model imagines and what actually happens cannot be measured on a historian at all, because the counterfactual trajectory was never run. A what-if study on the same flotation record is reported as a null with teeth: on the claims where the model responds at all it is inverted on every anchor, agreement 0.014 against a placebo of 0.526, because an operator raises the amine dose because the silica assay came back high. The record is confounded, the model reproduces the confounding faithfully, and no causal claim rests on it.