DreamerBuild: recorded Real Madrid–Barcelona experiment Recorded 15 November 2026. 3,000 simulated matches, not real football results. FILES inputs.json: complete team/lineup/bench snapshots, settings, seed root, recording time, source commit and source/runner/team-payload SHA-256 hashes. matches.csv: one row per simulated match, 1,000 for each venue. summary.json: counts, rates, averages, five common scores and Wilson intervals. checksums.json: SHA-256 of the input, match and summary files. TEAM ORDER AND SETTINGS Madrid is always the engine's home/first team; Barcelona is away/second. venue=home favors Madrid; venue=away favors Barcelona; neutral favors neither. All Home/Away statistical columns retain that team order, including away venue. Both starting formations: 4-2-3-1. Same recorded XIs and bench pools each match. Both initial tactics: balanced mentality, medium pressing, balanced tempo. Clear weather. Injuries on. Automatic substitutions on. Extra time/shootout off. The model may make substitutions and generate cards/injuries during a match. The dataset is a frozen tool snapshot, not verified official matchday squads. CSV COLUMNS venue: home/neutral/away, from Madrid's perspective. match: 1-based match index within a venue. seed: seedRoot|venue|match, the offline experiment's full RNG seed. engineSeed: additional engine result seed; not sufficient alone to replay a match. madridGoals/barcelonaGoals: final 90-minute score, including regulation stoppage. possessionHome/Away: possession percentages. shotsHome/Away, shotsOnTargetHome/Away: shot counts. xgHome/Away: simulated expected goals, not externally observed real-match xG. passAccuracyHome/Away: completed-pass percentages. cornersHome/Away, foulsHome/Away: counts. yellowCardsHome/Away, redCardsHome/Away: counts. METHOD Each seed is passed to the engine's createSeededRandom generator. During the synchronous simulateMatch call, all Math.random calls use that generator. The original Math.random is restored in a finally block after every match. Every match receives cloned inputs. No production engine code was changed. Venue groups use distinct seeds; these are not paired counterfactual replays. The seed was fixed before inspecting the output; this run was not selected from multiple runs to produce an attractive result. Counts and arithmetic means are calculated directly from matches.csv rows. Win-rate 95% Wilson intervals use z=1.96, with each venue treated separately. They quantify Monte Carlo uncertainty within this fixed model, not real-world forecast error. Small differences should not be read as conclusive superiority. REPRODUCE Use the DreamerBuild project checkout and the source commit recorded in inputs.json, with the experiment runner scripts/run-clasico-experiment.cjs from this article's implementation and these recorded files retained. Install the project dependencies with npm ci, then from the project root run: node scripts/run-clasico-experiment.cjs --verify The runner checks the bundled engine/lineup SHA-256, reruns all 3,000 matches, recomputes the summary and compares the CSV/summary byte for byte. A changed engine bundle is rejected. Node/esbuild versions are recorded in inputs.json. The source commit predates adding the experiment artifacts/runner; it identifies the engine and lineup code, not the commit containing this new article. Running without --verify creates the initial experiment and refuses to overwrite an existing inputs.json. Existing earlier blog experiments are not modified or validated by this new run.