Abstract
<title>Abstract</title> <p>Prior proposals for AI-assisted historical “what-if” simulation have argued qualitatively for the educational value of AI-generated counterfactuals, but none specifies a computational architecture, a causal-consistency mechanism, or an evaluation protocol capable of supporting scientific — as opposed to purely narrative — claims. This paper critically extends that line of work. We identify five concrete methodological gaps in existing AI-for-counterfactual-history proposals, situating them for the first time against the causal-inference, temporal-knowledge-graph, multi-agent-simulation, and quantum-optimization literatures. We then formalize historical counterfactual simulation as a constrained structural-causal-model (SCM) problem over a temporal knowledge graph, and propose CHRONOS — an original hybrid architecture integrating temporal knowledge graphs, Pearl-style do-calculus, retrieval-augmented multi-agent LLM simulation with causal-graph-constrained action legality, explicit uncertainty propagation, and a conditionally-invoked quantum/classical combinatorial sub-solver. We implement and run the core propagation and optimization mechanisms directly — not only describe them — on synthetic toy instances (a 9-node causal graph and a 6-qubit MaxCut sub-problem), and additionally report a scaled computational check (30 nodes; 12 qubits) that surfaces two honest, unflattering findings: naively hand-written consistency rules do not scale gracefully (contradiction rate rose from ~ 30% to ~ 90%), and QAOA’s approximation ratio is non-monotonic in circuit depth, consistent with known barren-plateau dynamics. We design an evaluation framework for causal consistency, calibration, and expert agreement; we are explicit throughout about which claims are established in the cited literature, which are reasoned extrapolation, and which are original, speculative proposals; and we do not claim, nor attempt to fabricate, the real-world empirical validation (a populated historical knowledge graph, a live multi-agent deployment, and independent domain-historian review) that a publication-ready version of this system would ultimately require.</p>