What does real-world, high-stakes decision-making require? The question is increasingly urgent, as we are rapidly offloading consequential decisions to “agentic” AI systems engineered around large language models (LLMs). The quality of those decisions is bounded by the capabilities of the systems sitting underneath them, and their cost is set by how those capabilities are achieved. Today the dominant answer is scale: more powerful models, longer reasoning traces, larger contexts, more agents reviewing other agents. The result is a cost/capability tradeoff that defines the practical application frontier of general-purpose LLMs. One might accept this tradeoff as the natural price of good decisions, but the history of prediction-intensive practices suggests an alternative path forward.
Consider how the practice of building a physical structure has matured. Building has one fundamental requirement, simply stated: the structure must stand. Whether a builder will succeed at this task depends in a highly specific way on their skill at predicting what will happen to the structure, during the building process and thereafter during its lifetime, under a wide variety of conditions and variations—a vast, implicit chain of “what-ifs”, many of which cannot be resolved definitively in advance, resulting in an extraordinarily peaked loss function. A child devises experiments using folk physics and builds by trial and error—enough for a tower of blocks, hopeless for anything long-standing. A medieval master builder drew on generations of accumulated heuristics: this affords far more complexity, but little adaptivity once circumstances depart from precedent. A structural engineer in the pre-computer era solved the governing equations by hand—genuinely predictive, but only as far as one expert can hold the procedural knowledge in their head: which equations to solve, which parameters matter, how to estimate them. And today, a builder who declines to use a CAD system—pre-loaded with scientifically validated models and parameters for every context, able to simulate a design in instants—is committing malpractice. And a “vibe-building” LLM-based system that purported to help with construction by retrieving heuristics and “reasoning through” the problem with mental math, however fast, would be considered an impressive demo of LLM capabilities, but for real work, such a system would necessarily have to offload the critical physical computations to a CAD engine.
The application of AI is quickly moving from simple, point-in-time decisions under extensive human supervision to autonomous, temporally-extended professional decisions, many of which interact with and impact partially-observable real-world systems, with running costs and serious risks. Yet in many such domains—investing, policy, operations—the state of the art still consists of accumulated heuristics and hand-solved models, held in expert heads. So long as humans hold responsibility and accountability for those decisions, they buffer against error with padding and oversight; but as their judgment is automated away by LLM-powered workflows in the name of efficiency, those buffers go with it. Delivering on the efficiency promise then requires a quality bar high enough to prevent catastrophic failures, at an operating cost low enough to justify the ROI.
Indeed, quality shortfalls are the largest driver of cost: an ungrounded prediction that is wrong—or merely unauditable—triggers verification and rework that compound downstream, so the true cost of a cheap answer is dominated by what happens after it is emitted. Conversely, the prevailing remedy for quality is to spend more inference, driving cost per decision up the frontier. For a one-off, low-stakes question, this bargain may be acceptable; but professional decision-making consists of repeated, high-stakes, evidence-coupled predictions, and there the tradeoff bites hardest exactly where it can least be afforded. This raises the question of whether the tradeoff is fundamental, or an artifact of the model class being asked to do the predicting.
We propose that the key capability required in these settings is structural causal prediction: (i) simulating the system forward under assumptions or hypothetical interventions (“what happens to this company’s free cash flow if the data-center build-out slows?”) and (ii) revising those forward simulations coherently as evidence arrives (“a supplier just guided down; update everything downstream”). Simulate-forward and condition-on-evidence are the defining operations of a causal generative model— what the CAD system does for the builder and the engineer-by-hand could only approximate.
We note that this capability is quite distinct from fluency, coherent-sounding reasoning, long-running memory, theoretical knowledge, mathematical problem-solving prowess, or any of the other capabilities identified heretofore in LLMs. An LLM can describe such a model in prose, and with a coding tool or spreadsheet it can even assemble and run an ad hoc one; what it cannot do is be one, because its outputs are not constrained to be the posterior of any explicit mechanism.222Indeed, even if an LLM’s output momentarily coincided with the right posterior, it could not remain one as evidence arrives, because its update operator—accumulating context or re-querying a knowledge base—is not Bayes’ rule (Proposition 3). A reliably grounded result therefore requires the LLM to call an explicit, verified model rather than to be one. The moment an agent does build, ground, verify, and update such a model, it has stopped doing free-form language modeling and started constructing the grounded world model we formalize here. This is consonant with Garcez (2025)’s argument that context augmentation and post-hoc RLHF cannot fix hallucination, since errors compound in continuous, ungrounded computation regardless of test-time compute, and that reliability instead requires coupling the network to an explicit, extractable symbolic mechanism.
Motivated by the above, we define a grounded world model (GWM) as a causal generative model of a domain whose structure mirrors the domain’s actual entities and mechanisms, whose parameters are grounded—tied to evidence by explicit, auditable inference—and which supports the two operations above natively: it can be intervened upon (simulate-forward) and conditioned on incoming evidence (condition-on-evidence) to return a calibrated posterior. What makes a model a GWM is this functional contract, not any particular implementation. A probabilistic program is one natural realization, and the one we use; but a GWM may equally be expressed as a structural causal model, a system of dynamical equations, or a simulator equipped with a likelihood—any representation that supports verified intervention and conditioning qualifies.
Fully-fledged GWMs already exist. Numerical weather prediction replaced expert pattern-matching with forward simulation of a grounded physical model, continuously corrected by assimilating observations—the “quiet revolution” (Bauer et al., 2015) that has gained roughly a day of forecast lead time per decade. We believe the same move can be brought to other domains, particularly business and finance; this belief motivated Primordia’s GWM for equity investing.
The difference in aims between GWMs and LLMs—GWMs directly predict real-world observables; LLMs predict token sequences in response to a decision-maker’s prompts, which can then be interpreted by her as predictions of the world—is reflected in architectural and engineering differences. Most strikingly, LLMs are generically underdetermined (hence the “L”), monolithic, dense, static, and stateless, while GWMs are parsimonious, modular, sparse, dynamic, and stateful (they update as the world changes). A GWM’s development and maintenance process is therefore rather different than an LLM’s, especially when (as in our case) complete mechanistic explainability is required. Primordia’s approach is based on a combination of continual, automated program synthesis, continual validation of predictions against outcomes, and expert-informed structural priors.333We make no claim to our process being unique, nor claim optimality. Notably, alternative approaches are well-known in the literature, such as ones based on causal discovery (Spirtes et al., 2000), evolutionary methods (Koza, 1992), and symbolic regression (Schmidt and Lipson, 2009). A full comparison between processes is beyond the scope of this paper.
This paper aims to formalize the category of GWMs, to demonstrate that they dominate “LLMs-as-predictors” jointly on cost and quality—escaping, rather than optimizing along, the LLM cost/capability tradeoff—and to benchmark Primordia’s GWM implementation against a typical “LLM-as-predictor” workflow that a principal would otherwise deploy. We draw the running example from investment analysis, where Primordia offers a GWM-based platform since December 2025; but the argument is about the model class, and we flag its domain-independence throughout.
Structural causal prediction requires a GWM. We define the task category and show that it requires a GWM (Section 2). LLMs and knowledge graphs fail to meet the requirements for structural reasons, not for want of scale—indeed, the prior probability that a language model’s circuit happens to implement exact conditioning decreases with parameter count (Proposition 3, Corollary 2).
LLM predictors’ explanation quality has a structural ceiling. A prediction is only as useful as the explanation behind it. We operationalize explanation quality (XQ) from observable, mechanism-checkable properties and prove it lower-bounds prediction quality (closeness to the ideal predictor), so a rational agent trusts a prediction in proportion to its XQ (Section 2.7, Proposition 1). An LLM predictor’s XQ has a ceiling strictly below the GWM’s, and no added inference budget can clear it, whereas the GWM sits at its conditioning-limited ceiling at flat cost in the quality target (Section 3).
GWM prediction is cheap, quickly amortizing build costs. Because a GWM answers each query by inference on a structure-aligned model rather than by a fresh narration, it is both far smaller and far cheaper per prediction than an LLM, and its one-time build cost is recovered within a handful of predictive distributions (Section 3).
LeCun (2022) argues that autonomous intelligence requires a learned world model supporting prediction and planning. We sharpen “world model” to the verified, parametric, causally-transparent causal model required by structural causal prediction, and contrast it with the implicit, unverified world model latent in an LLM. The GWM is the “symbolic” half of the neurosymbolic architecture in Garcez and Lamb (2023); Garcez (2025), with the LLM as the “neural” half; more broadly, Chiatti et al. (2026) argue that systematic neural/symbolic integration, not scale alone, is the route to reliable AI in critical domains, and the GWM is one concrete, deployed instance. Closest in spirit is Wong et al. (2023), who translate natural language into probabilistic programs (a “probabilistic language of thought”); we take the further step of grounding, verifying, and composing such programs into reusable foundation-model capital. This paper builds directly on our own prior work, Kaufmann et al. (2025), which introduced the pre-synthesized grounded world model—a synthesized, continually-refined probabilistic program supporting full Bayesian inference—and showed it reaching Bayes-optimal accuracy on a synthetic-equities benchmark where SOTA LLMs plateau near 40%; here we formalize the GWM as a model class and benchmark its cost and quality against narrative-world-model baselines. The GAIA tech tree (Walters et al., 2025b) instantiates the same neural/symbolic split—the LLM exploring, the verified model library serving as the symbolic half it retrieves from—for the domain of technology planning. Closely related in stance is the Scientist AI program of Bengio et al. (2025a): a non-agentic world model that generates theories to explain data, paired with an inference machine carrying explicit uncertainty, proposed as a run-time safety guardrail whose Bayesian risk bounds can reject dangerous actions (Bengio et al., 2025b). A GWM is a domain-specific, verified realization of exactly this “model that explains rather than acts” posture. Finally, using a GWM for control (Section 4.1) is active inference (Friston et al., 2017): acting to minimize expected free energy under a generative model.
The GWM can be seen as a structural causal model (Pearl, 2009), which we represent as a probabilistic program (Koller and Friedman, 2009). That a causal model is necessary, not merely convenient, is established by Richens and Everitt (2024): any agent satisfying a regret bound under a large set of distributional shifts must have learned an approximate causal model, converging to the true model for optimal agents. We take this as the formal warrant for selecting the GWM by requirements.
A rapidly expanding family of systems equips an LLM with external structured knowledge or capabilities: knowledge-graph question answering (Baek et al., 2023; Edge et al., 2024); long-context retrieval (Lewis et al., 2020); “agentic memory” stores that accumulate and re-read facts across calls (Packer et al., 2023); and tool-augmented agents that call a calculator, spreadsheet, or code interpreter to compute rather than merely retrieve (Schick et al., 2023; Yao et al., 2022; Gao et al., 2023). The retrieval and memory members of this family share a ceiling regardless of sophistication: absent an explicit mechanism, they improve only the evidence an LLM can cite, so the model still narrates a forecast rather than computing one, gaining neither the grounding and auditability of a GWM nor proper Bayesian updating as evidence arrives. Consistent with this, Vals AI (2026) benchmarks tool-augmented SOTA LLMs on analyst-grade financial research and finds they plateau well short of reliability, even given unlimited test-time budgets.444As of writing this, the best-performing LLM (Gemini Flash 3.5) scores below 58%.
A computation tool (i.e., code execution) is different in kind. Raising explanation quality with such a tool means using it to construct, on the fly, the very object we formalize: an explicit mechanism the LLM can condition on and re-query, i.e. an ad hoc GWM assembled at inference time. This is not a counterexample to our thesis but an instance of it: the rational endpoint of tool-augmentation, pushed far enough, is exactly the Mode-2/Mode-3 coupling of Section 4.1, where the LLM frames and explains while a verified model computes the posterior—a per-query, un-amortized version of the architecture we propose building once and reusing. Tellingly, the harness in (Vals AI, 2026) does not supply code-execution tools, foreclosing that path by construction.
Two neural-side programs each share one property of a GWM while lacking the others. Physics-informed machine learning (Karniadakis et al., 2021) shares the commitment to mechanism, regularizing a network with known governing equations; but the result is a neural surrogate whose outputs remain unverified and whose evidence-absorption is not Bayesian conditioning, whereas a GWM keeps the mechanism an explicit program and its updates exact. Post-hoc explainability instead shares the aim of explanation: a large literature extracts accounts from a trained black box, ranging from local attribution (Ribeiro et al., 2016; Lundberg and Lee, 2017) to global, model-level counterfactuals (Sobieski and Biecek, 2024); a GWM inverts this, since its explanation is the mechanism and requires no extraction (Section 2.7).
Farmer (2024) argues that economics should make the move meteorology made—from equilibrium and reduced-form statistics to mechanistic, agent-level simulation—and conjectures that economic systems may be more tractable than the weather. We operationalize that program for structural causal prediction and take up the tractability conjecture in Section 4.
Section 2 builds a taxonomy of world models, defines the GWM and its grounding map, fixes its defining properties, positions it against alternatives, and formalizes prediction and explanation quality (proofs in Appendix G). Section 3 presents benchmark results, pricing the cost of a single grounded prediction, the cost to reach a target explanation quality, and the volume-and-lifetime costs. Section 4 consolidates the foundation-model, leverage, parsimony, and usage-mode discussion together with the groundedness ladder, generalization and the tractability hypothesis, and limitations; Section 5 concludes. Appendices define the explanation-quality components (Appendix C), work a single case end to end (Appendix E), collect the cost-model details (Appendix F) and the proofs including the XQ class ceilings and GWM dominance (Appendix G), and tabulate the symbol, value, and provenance of every numeric input (Appendix I).