Carapace Intelligence

Carapace Intelligence is decision-intelligence infrastructure for organizations operating in adversarial, multi-party environments. It treats an organization's situation as a finite directed graph of states and actions, uses LLMs as local judgment elicitors (transition probabilities, outcome scores) rather than global reasoners, and then runs explicit Bayesian expectimax search and sensitivity analysis on top. The result is a ranked set of recommended actions with quantified robustness - an advisor that can surface non-obvious moves and explain why they hold up under perturbation.

The work spans both the theory and the implementation. Two formal papers landed in May 2026: one specifying the full mathematical pipeline (finite DAG, per-state value functions, Bayesian expectimax backward induction, multi-arm sensitivity analysis), and one characterizing why these games sit outside the reach of classical game theory. Alongside them, a deterministic Python pipeline implements the paper end to end - frozen-dataclass types, a content-hashed rationale cache, the solver, sensitivity sweep, self-validating elicitation agents, tree construction, a tail-risk audit gate, and a CLI - with the LLM confined to local judgments so all global reasoning stays reproducible from cached outputs.