Representative Publications:

• Elin, S.M. (2025). The mathematical machinery of causal inference: from data to decision advantage.

  Intelligence and National Security, 1-31.https://doi.org/10.1080/02684527.2025.2574789

  Abstract: Quantitative intelligence analysis often leans on pattern recognition, but in adversary‑shaped environments correlations can be engineered. Building on Judea Pearl’s structural causal models, this article is makes identification, not estimation, the gate to credible claims. It shows when effects are recoverable from observational data and when leverage—mediators, instruments, modest interventions, or transport adjustments—is required. The framework unifies association, intervention, and counterfactuals and extends to sequential and multi‑agent settings (Elias Bareinboim’s causal reinforcement learning and causal game theory), where strategies are modelled as interventions and evaluated under adaptation. Across operational cases, analysts specify a causal graph, test its implications, determine identifiability, and only then estimate. This identification‑first discipline separates artefacts from effects, hardens AI/ML pipelines against manipulation, and ties analysis directly to ‘decision advantage’: identify → estimate → decide → iterate. The result is an explicit, testable mapping from action to outcome that turns uncertain signals into reliable policy support. Echoing Sherman Kent, intelligence analysis earns trust when it distinguishes observation from inference and states what is known—and unknown—up front.

Forthcoming & Working papers:

• Elin, S. M. (2027). Causal reasoning in open-source intelligence: inference, intervention, and counterfactuals for decision advantage. In R. Shaffer (Ed.), Handbook of Open-Source Intelligence. TBA.

  Abstract: The expansion of open-source intelligence (OSINT) has widened access to public data without resolving the central problem of intelligence analysis: how to turn observation into dependable judgment. In information-rich and adversarial environments, digital traces, public signals, and visible events may inform analysis, but they may also reflect deception, selection effects, and non-causal dependence. The difficulty is therefore not only one of scale or access, but of inference. More data can improve analysis, but it can also amplify noise, reward superficial pattern recognition, and encourage unwarranted movement from correlation to explanation. Recent work on OSINT has usefully emphasized that the field is better understood as an evolving set of practices for exploiting open sources than as a revolutionary break with earlier intelligence methods. This chapter argues that OSINT should be understood not merely as a means of information acquisition, but as a form of inquiry that requires explicit causal reasoning. Drawing on causal inference, interventionist theories of explanation, and adjacent work in formal epistemology, machine learning, and the decision sciences, it outlines a framework for moving from observed regularities to claims about mechanism, intervention, and counterfactual dependence. The aim is not to impose abstract formalism on intelligence practice, but to clarify what kinds of conclusions open-source evidence can support, under what assumptions, and with what degree of confidence. In this respect, the chapter extends a central lesson of recent causal work in intelligence analysis: identification must precede estimation, especially where actors may shape the observable environment in ways designed to mislead inference. The chapter shows how causal reasoning helps distinguish descriptive from explanatory claims, identify when available evidence supports causal judgment, and determine when stronger assumptions, additional data, or experimental leverage are required. Under such conditions, the task of OSINT is not only to collect and verify open information, but also to reason carefully about the processes that generate it. The result is a conception of OSINT oriented not simply toward acquisition, but toward explanation, evaluation, and judgment under uncertainty.

• Elin, S.M. (2026). Knowing in Hell: Machiavellian Realism and Resource Rationality.

  Abstract: This paper proposes a resource-rational reconstruction of Machiavellian realism. I model political judgment as a problem of bounded causal inference in environments characterized by conflict, strategic opacity, and costly information. Political agents do not deliberate under conditions of epistemic transparency; they act under pressure, with limited computational resources, while other agents actively manipulate signals, conceal structure, and respond strategically to observation and intervention. I argue that resource rationality provides a general framework for analyzing such settings because it explains how finite reasoners optimally allocate attention, search, and inferential effort relative to the structure of the task environment. Causal reasoning supplies a complementary vocabulary for intervention, counterfactual dependence, and manipulable structure, but must itself be embedded in a model of uncertainty, cost, and adversarial interaction. On this view, political knowledge is not well described as approximate ideal deliberation. It is better understood as a form of adaptive, resource-bounded, strategically situated causal judgment. This account clarifies the epistemic core of Machiavellian prudence and suggests a more general theory of realism under cognitive constraint.

• Elin, S.M. (2026). Equilibria Under Intervention: AI Safety in Strategic Multi-Agent Systems.

  Abstract: Many discussions of AI safety evaluate interventions at the level of an individual model. Typical questions are whether a training procedure reduces harmful behavior, whether an oversight mechanism increases compliance, or whether a restriction on model access limits misuse. These are useful questions, but they are often posed in an implicitly single-agent setting. In practice, safety interventions are deployed in environments containing other adaptive actors, such as users, monitors, competing systems, organizations, and governments. In these settings, an intervention does not simply modify one component of the system. It also changes incentives, information, and feasible responses for other actors. For that reason, the effect of a safety intervention cannot be identified with its immediate effect on the target model. The appropriate object of analysis is the equilibrium induced by the intervention. This article argues that a multi-agent perspective is necessary for understanding when safety measures are robust, when they are offset by strategic adaptation, and when they shift rather than reduce risk. The main conclusion is that AI safety in strategic settings is best framed as the design and evaluation of equilibria under intervention.