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

Version 0.01. 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 preprint is about. We describe latentplant, an open (MIT, PyPI) engine for action-conditioned latent dynamics on industrial process time series, and 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 (an hourly laboratory assay carried on twenty-second rows repeats about 180 times, and a model trained without masking those repeats learns to copy the previous row, scores beautifully, and knows nothing); a leakage-safe windowing layer; a ladder of rungs from persistence through a standardized ridge VARX and a PETS-class probabilistic ensemble to a recurrent state-space model and an ensembled RSSM; 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 all three are reported. On a one-dimensional Burger-Diehl-class thickener PDE whose entire solids concentration field is hidden behind four instruments, the recurrent latent attains the best point accuracy of any rung (RMSE 0.1680 against 0.2528 for the next best and 0.2686 for persistence, a skill of +0.375 against persistence where every other rung manages about +0.059) and is simultaneously overconfident, its nominal 90 percent intervals covering 0.672 of outcomes, so a proper scoring rule ranks it third. Both facts are true at once, and that single measurement is why conformal correction is in the engine. On a grinding circuit that is smooth and near-linear about its operating point, a standardized ridge VARX wins outright, which is the expected result rather than a disappointment. On Tennessee Eastman the same engine wins clearly on fault-free dynamics, and its surprise stream detects 14 of 20 fault classes at a measured 1.00 percent per-sample false alarm rate fitted on fault-free runs alone, while all three known-undetectable leakage probes (IDV 3, 9, 15) correctly go undetected, which is what makes the other fourteen credible. On a real iron-ore flotation record the world model does not beat holding the last value at either cadence tested. The minute-cadence case had declared, before running, that a second failure would reject the hypothesis that hourly averaging destroyed the fast response; at sixty times the temporal resolution and 245,820 rows it failed again, so that explanation is rejected. The mechanism was then measured rather than assumed: the principal exogenous disturbance, feed grade, has a median run length of nine hours, so within a sixty-minute horizon it does not move at all, while the actions were set by a closed loop responding to a disturbance the model cannot see changing. This is a limit of observational plant data rather than of the method, and reporting it is the most useful contribution here: a field that reports only its wins cannot distinguish a method that works from a dataset that was easy. The preprint also argues why the two reference environments ship with the engine rather than with a consuming application: 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. Every numerical value in the manuscript is generated from committed artifacts rather than transcribed, and the paper states its refusals structurally: no causal identification from logged data, no setpoint advice, no confidence intervals over overlapping windows, and no comparison against a mineral-processing world-model benchmark because none exists. Engine: https://pypi.org/project/latentplant/ (MIT). Source, reference environments, declared kill criteria and physics tests: https://github.com/fsantibanezleal/CAOS_LatentPlant .