Counterfactual Evaluation by Perturbation of Information States: Deleting the Ordering, Not Refining It

We propose a semantics for counterfactuals on which evaluation never leaves the evaluator: to assess if X had happened, Y would have happened, one perturbs the evaluator’s own information state — reopen the antecedent’s entry at a declared scope, restate it to the antecedent’s value, re-prune by the state’s consistency constraints — and reads the verdict off the surviving residue. Would, might, and the probability that are one operator read at three grains over that residue. The proposal’s headline is a deletion with an address: the similarity ordering of the Lewis–Stalnaker tradition appears nowhere in the clauses, and its work is identified as split among three independently needed structures — the dependency structure of the state, binary validity, and graded weight. The same deletion lands on the ordering source of the standard natural-language treatment of modality, while the modal base survives as the residue. We run the tradition’s own adversarial battery (the failed inference patterns, Bizet–Verdi, backtracking) and add a case no whole-world ordering supplies: horizon-indexed graded counterfactuals for chaotic systems. We state exactly what the semantics is committed to — most importantly, that learning must be modeled as transit to a new state rather than revision in place, on pain of the indexed verdicts collapsing — and we pose the proposal’s standing formal challenge: whether any ordering semantics can generate the indexed, weighted, multi-pocket verdict function across evaluators, the equivalence results of Lewis and Grove notwithstanding.