Mohsen Mostafa · Preprints.org 2026 · 2026
DOI: 10.20944/preprints202610.1780.v1
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Guarantees for latent world models have largely concerned representation quality—decodability, conditioning, isotropy—rather than whether a plan produced by the model is close to what the environment would actually do. This paper formalizes five conditions that any control-sufficient latent world model must satisfy and proves conditional theorems with explicit vacuity thresholds: a control-sufficiency-to-planning-regret bound, a finite-sample exploitability-calibrated gate, a Mondrian conformal certificate under contact, a contact-mode identity, and a planner-feasibility upper confidence bound. Under the measured premises of the two prior papers in this program, Dynamic LeJEPA [1] (maximum-entropy encoders with physics in the decoder) and DEGP [2] (verified conditioning with 0/50 deployment), we prove conditional lower bounds showing that several of the corresponding vacuity parameters cannot be small in that setup. The bounds are conditional; they do not claim general impossibility. We then report a 30-seed replication on a preserved copy of the standard Push-T corpus: 180 encoder–predictor pairs across six ranks, 150 control models (120 attribution controls + 30 baseline re-runs), and 90 trained decoders. Three measured findings replace earlier readings. Contact selectivity is flat: it peaks at 1.04× at r = 128 against a 2.0× target (p = 0.96 versus r = 0). Shuffled-target and random-subspace controls are indistinguishable from baseline at every rank. Free-stratum skill is negative at every coupled rank. Decoder cross-episode position error is 40.7–42.2 px—an improvement over the prior ≈ 50 px but far from the < 20 px required for a usable task tolerance. Split-conformal certificate widths at H ∈ {4, 8, 16} are 126–279 px at nominal 0.90 coverage on random episode splits and 0.862–0.914 on a chronological split. A pre-specified oracle-feasibility prediction (P4: coverage-shaped cost raises poracle ≥ 0.5) is evaluated and falsified: 2/50, versus 13/50 for the terminal-pose cost on the same seeds. The contribution is a boundary: a map of the exact assumptions a control-sufficient latent world model requires, each lower bound instantiated on measured data, and a pre-specified falsification protocol (P1–P6) that a solution must pass. The paper claims no solution and no state-of-the-art result.
No comments yet — start the discussion below.