{
 "approach": "Frozen paired measurement of candidate side-jump-budget (per-side marginal jump charge at main bid/ask sizing) against current best B (combine-cash) on confirm block seeds 256000..256999; no code changed.",
 "tried": "Downloaded both artifact JSONs, extracted the `code` field, verified SHA-256 of UTF-8 bytes before running (candidate 3b7236c9... and B baef955e... both matched). Simulator cloned from github.com/danrobinson/prediction-market-challenge, checked out b244530ee2a6247b5f930830eeca9dd950da0e7e, unmodified, Python 3.10.12. Each seed run in process via load_strategy_factory(path) and run_batch(factory, base_config=ChallengeConfig(), variance=ParameterVariance(), n_simulations=1, seed_start=seed, workers=1); default 2000 steps; no --workers, no --sandbox. Timed 10 seeds per strategy first (~0.76 s/sim). Block run as 8 OS processes (4 at a time) on disjoint seed ranges of 250, each strategy on every seed 256000..256999; ~6.5 min wall. Checked every seed present for both, no simulation raised or returned failed=True, no other seeds used beyond the 10-seed timing run inside the block (256000..256009, rerun as part of the block). Candidate wins on 349 seeds, loses on 646, ties 5. Means: candidate retail 126.1113 arb -24.5097; B retail 127.8088 arb -26.6107.",
 "next": "On this block the candidate's gain (+0.404, se 0.499) is under one standard error: not a promotion signal. It trades retail edge (-1.698 +/- 0.168) for lower arbitrage loss (+2.101 +/- 0.472); both component shifts are significant but nearly cancel. Pool with the other r2-confirm blocks before any finalist run; a version that recovers retail edge while keeping the arb saving (e.g. charging only on the side the jump would hit) may be worth a development item.",
 "code_sha256": "3b7236c98d52efed9617e4cf6d8183012c98b8b61b18eb59b5afae937c99c724",
 "incumbent_code_sha256": "baef955ec10728839f7c3357016ec9eb0550e7ec689bbd5171b1cdd5c9e37273",
 "tuning": null,
 "paired": {
  "seeds": "256000..256999",
  "n": 1000,
  "complete": true,
  "mean_diff": 0.403519,
  "se": 0.499107,
  "sum_d": 403.519188,
  "sum_d2": 249021.102873,
  "candidate_mean": 101.601627,
  "incumbent_mean": 101.198108,
  "retail_diff": -1.697523,
  "retail_se": 0.168357,
  "arb_diff": 2.101043,
  "arb_se": 0.472233,
  "nonzero_diff_seeds": 995,
  "failures": []
 },
 "seconds_per_simulation": 0.763
}