What a Grenade Is For: An Annotation Schema and a Pre-registered Human-LLM Test for Utility Lineups in Professional Counter-Strike

Abstract

Grenade use in professional Counter-Strike follows community conventions written neither in the game's rules nor in its physics. We present an annotation schema for utility lineups, defined by an origin zone, a landing zone and a grenade type, and a de_dust2 lineup dataset built from 69,827 throws in 122 professional matches of the ESTA dataset. Four fields are computed without human input; three of them were tested against criteria registered before they were derived. Geometry and frequency are usable with stated limits. Generality met its registered criterion for being unusable (rho_MM = -0.217, p = 7.1e-5, 330 lineups) and is not used. Difficulty passed its gate on bucket count but failed its main test: the tier gap was -0.98 with a 95% cluster bootstrap interval of [-6.68, 2.94], so it is reported only as a description. The purpose of a throw (intent) and whether it is conventional are left to human annotators. We pre-register a test of whether a panel of large language models could replace them. Three experienced players and at least three vision-capable models will label the same 100 stimuli under identical instructions. Krippendorff's alpha with item-bootstrap intervals will be judged against thresholds of 0.5 and 0.7 through five branches with a fixed verdict order. We report the full registration history, including revisions that did not only tighten the plan and an AI-filled label file that was never used. No human label exists yet. Results will appear in a later version.