Frozen-opponent replay and a paired live control in Kaggriculture: code, per-game results, episode identifiers and dated plans

Code and data behind a study of frozen-opponent replay in Kaggriculture, a Kaggle simulation competition (two-player farming with a shared market; 10,148 teams; evaluation ended 14 October 2026). The study replays candidate bots into recorded ladder games with the opponent frozen to its recorded actions, on the official engine with a price correction. The instrument reproduced the deployed bot's games exactly and predicted a gain of 16.0 percentage points [9.7, 22.3] for an opponent-conditioned portfolio on the games it targets. A paired control fixed in advance, run on all 875 usable live games of that portfolio, found -0.5 points [-5.3, +4.4] on the 627 targeted games. An unpaired check that uses no replay, against an unchanged copy of the baseline playing real games over the same window, gave -0.2 points [-5.7, +5.3] on targeted games. A positive control shows the same design detects a real gain of +13.5 points [+10.2, +16.9]. Contents: the counterfactual harness and analysis scripts; per-game results of every replay and live analysis used; the episode ids and start times of every live game; the submitted bots (Apache-2.0, upstream notices kept); the dated analysis plans and verbatim log excerpts; the final leaderboard as ranks and scores without team names; reproduce_live.py, which recomputes the live control table from the per-game file, and paper_numbers.py, which rechecks every number in the manuscript that these files allow. Opponent team names are replaced by stable pseudonyms. Game replays are not redistributed; they can be downloaded from Kaggle by episode id. The README gives the environment (Python 3.12, kaggle-environments 1.32.7), the exact commands, and known issues (a bug in one download script that delayed the real-game check, now fixed and documented).