Grounded World Models: Efficient and Verifiable Structural Causal Prediction

Abstract

Investment decisions require predictive forecasts under hypothetical interventions which are coherently justified by the sum of available evidence. We argue this problem class selects a model class: the grounded world model (GWM), a verified, composable causal model whose every prediction is a Bayesian posterior under an explicit mechanism. We formalize GWMs and compare them theoretically and empirically with the default alternative: directly prompting a large language model (LLM) for predictions. Theoretically, we establish dominance results along both the efficiency and quality axes, which compound. Empirically, we benchmark Primordia’s investment GWM against a SOTA LLM (Claude Opus 4.8) on a standard investment analysis task. We show that GWMs’ explicit mechanistic, theory-aligned structure allows for $\sim {10}^5\times$ parsimony compared to general-purpose LLMs, enabling incremental prediction at $\sim 36,000\times$ lower cost and $\sim 350\times$ lower latency, while achieving $\sim 2.4\times$ greater quality. Further, we show that any LLM predictor’s quality stays below the GWM’s at any finite cost, with cost per prediction diverging as that shared ceiling is approached. We propose a hybrid neurosymbolic structure—GWM as predictor, LLM as natural-language interface and query orchestrator—as the theoretically and empirically superior architecture for investment and other high-stakes real-world decision domains.

Figure 1: From narration to a grounded posterior. An LLM (even with a knowledge graph) answers a query with a prediction $q(Y)$ that is not constrained to be any posterior—ungrounded and not auditable. A GWM casts a prediction as a proper Bayesian posterior under a causal generative model—grounded, source-traceable, and cheap.