How do we evaluate the quality of a prediction when the ground truth has not yet occurred? If an artificial intelligence forecasts the price of Bitcoin in 2030 or the outcome of an election years away, standard accuracy metrics are useless in the present.
The key insight behind a long-standing and rapidly revitalizing research direction is that we do not need to wait for the future to evaluate probabilistic beliefs. Instead, we can audit them for internal coherence. Probability theory is not merely a set of loose guidelines; it is a rigid system of mathematical axioms. If an agent's assigned probabilities violate these axioms—for instance, by claiming a specific joint event is more likely than its individual components—the system is logically broken from within. By evaluating and enforcing these axiomatic boundaries, we can validate, benchmark, and even repair probabilistic reasoning systems entirely "outcome-free."
The Geometry of Coherence
To understand why outcome-free auditing works, we must look at the mathematical constraints that probability theory imposes on logical propositions.
Consider two statements, $A$ and $B$. You are free to assign any probability you like to $A$ happening, and any probability to $B$. However, the moment those two values are set, the probability of their conjunction, $P(A \land B)$, is no longer completely free. It is strictly bounded by the laws of probability:
$$\max(0, P(A) + P(B) - 1) \le P(A \land B) \le \min(P(A), P(B))$$
If a system assigns a value to $P(A \land B)$ that violates these bounds, it has committed an axiomatic error.
This geometric reality was formalized early in AI research. An early foundational paper [75] established that ensuring a consistent probability assignment for $k$ basic propositions requires satisfying a system of equations over $2^k - 1$ a priori probabilities, proving its equivalence to classical probability spaces.
Guggenheimer et al. [Foundations of Probabilistic Logic] built upon this to derive the necessary and sufficient conditions for consistency, demonstrating that early expert systems' "certainty measures" frequently broke basic logical bounds, such as the conjunction minimum.
When integrating knowledge from multiple distinct sources, maintaining this alignment becomes a complex linear programming problem. As Gilio et al. [Knowledge integration for Conditional Probability Assessments] demonstrated, simply checking the compatibility of derived linear systems is a necessary but not always sufficient condition to guarantee that combined conditional probabilities remain coherent and free of logical contradictions.
The Coherence Deficit in Language Models
While these geometric rules were originally applied to expert systems and databases—such as the "pseudoconsistency" checks designed by Lu et al. [Probabilistic object bases] to verify uncertain object schemas—they have taken on urgent relevance in the era of Large Language Models (LLMs).
LLMs are frequently deployed as predictive engines, yet they generate text via statistical next-token approximations rather than explicit probabilistic calculus. As a result, they frequently suffer from a severe coherence deficit.
Toni et al. [Exploring the Potential for Large Language Models to Demonstrate Rational Probabilistic Beliefs] exposed this by benchmarking state-of-the-art models on structured claim sets containing base statements, negations, and hierarchical variations. They found that models routinely violate basic axioms like complementarity ($P(A) + P(\neg A) = 1$) and monotonicity (assigning a higher probability to a specific subset event than its broader superset).
To turn this insight into an evaluation tool, Paleka et al. [Consistency Checks for Language Model Forecasters] introduced a framework that evaluates forecasting models by querying them on logically linked question tuples. By testing whether an LLM's predictions adhere to the law of total probability or conditional independence across these tuples, they can measure a model's forecasting competence without needing a single real-world outcome.
Similarly, Yang et al. [LLM-as-a-Prophet: Understanding Predictive Intelligence with Prophet Arena] leverage subset-superset sanity checks within live forecasting environments to distinguish whether an LLM's predictive intelligence is grounded in coherent reasoning or superficial statistical mimicry.
This lack of coherence also undermines how we modify model knowledge. When we attempt to update a fact inside an LLM, we are performing belief revision. Hase et al. [Fundamental Problems With Model Editing: How Should Rational Belief Revision Work in LLMs?] argue that current model editing techniques are fundamentally flawed because they treat edits as isolated sentence corrections. By evaluating edits against a web of logically related beliefs, the authors show that standard edits break probability axioms across connected statements, failing the basic requirements of rational belief revision.
Outcome-Free Auditing and Repair
If an agent's probabilities are incoherent, they are vulnerable to "arbitrage"—meaning an opponent could construct a set of bets against the agent that guarantees the agent will lose money, regardless of how the future unfolds.
Can we fix this incoherence without knowing the true outcomes? Because the space of valid probability distributions forms a convex set (a simplex), any incoherent assignment lies outside this valid space. This geometric insight enables a powerful mechanism: we can mathematically project the incoherent values back onto the closest valid point on the simplex.
Incoherent Belief (Outside Simplex)
\
\ <- Mathematical Projection (Repair)
v
+-----------------------+
| Valid Coherent Space |
| (Simplex) |
| |
| * Repaired |
| Belief |
+-----------------------+
Li et al. [Outcome-Free Arbitrage Audits and Coherence Repairs for LLM Forecasters] operationalized this by designing a closed-form "coherence repair" system. They audit an LLM's predictions using a graph of logical constraints (complementary pairs, monotonicity chains, and conjunction bounds). When violations are found, the system solves a constrained optimization problem to project the predictions back into coherence. Crucially, the authors prove mathematically that this projection is guaranteed to improve or maintain the accuracy of the forecast, providing a true "free lunch" for predictive performance without requiring any external data.
For autonomous agents that must reason internally without generating paradoxes, this axiomatic discipline can be baked directly into the underlying logic. Delgrande et al. [Epistemic Logic of Likelihood and Belief] designed the Epistemic Logic of Likelihood (ELL) to enforce these boundaries natively. To avoid classic pitfalls like the Lottery Paradox (where believing every individual ticket will lose contradicts believing that some ticket will win), their framework dictates that the probability of every individual world within a belief set must strictly exceed the combined probability of all excluded worlds. This constraint ensures that the principle of conjunction holds perfectly, allowing agents to maintain stable, rational beliefs under uncertainty.
Ultimately, this research direction demonstrates that truth and internal consistency are deeply intertwined. By treating the laws of probability as absolute geometric boundaries, we can audit, evaluate, and correct artificial minds entirely from within.
Key Foundational Work
Knowledge integration for Conditional Probability Assessments (1992) — contributed 5 concept elements to this direction.
Foundations of Probabilistic Logic (1987) — contributed 3 concept elements to this direction.
Foundations of Probabilistic Logic (1987) — contributed 1 concept elements to this direction.