Before you use it: Validating a tactical command vocabulary for multi-agent shooters

Abstract

Multi-agent reinforcement learning evaluates communication by its effect on task reward, which cannot separate messages that were understood and acted on from messages that were ignored while the policy improved for other reasons. Separating them needs a verdict on the individual message, and that needs a discrete command space—whose own validity is not usually examined. We freeze a tactical command vocabulary for Counter-Strike (28 verbs, 7,936 legal messages, 6 compliance predicates) and, before using it, measure it against 270 professional matches: coverage, stated as the share of observed events the mechanics cannot produce; collapse, the structure surviving projection, swept over granularity rather than fixed; and predicate sharpness, evaluated counterfactually over each predicate's parameter space, needing no compliance labels. The action space expresses 87.7–89.2% of professional kills; symbolisation retains 1,305–1,533 effective utility usages and 31–32 reachable objectives, under saturation checks; the predicates accept 1.00–4.71 candidates over the map's real callout regions, giving chance floors of 3–25% that any reported compliance rate must be read against. Applied to the closest existing abstraction of this game, the same measurements corroborate two of its design claims and price a third. None of this requires an agent, and seven statements written into specifications and never checked were falsified along the way, three of them in the vocabulary file we release.