# Стенд-проверка (stage-2) · промт оператора: (универсальный) Ответ агента: { "worked": true, "confidence": 8, "verdict": "The sequence track is structurally matched to the attention intervention. The gated local-global transformer achieved lower test MSE than the tuned dense transformer (0.372641 vs 0.403156), with 7/8 paired wins and permutation p=0.01475, satisfying the harness criterion for a significant win. The trained-model gate signature confirmed use of both local and global paths.", "metrics": { "baseline": "Dense transformer_tiny, tuned lr=0.001, 8-seed mean MSE 0.403156, std 0.052733", "idea": "Gated local-global transformer, lr=0.001, 8-seed mean MSE 0.372641, std 0.068380", "delta_mean": -0.030515525490045547, "p_value": 0.01475 }, "bench_report": { "bench_version": 1, "track": "sequence", "model": "transformer_tiny", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.001 }, "sweep": [ { "cfg": { "lr": 0.001 }, "mean": 0.37249068170785904 }, { "cfg": { "lr": 0.003 }, "mean": 0.4498218819499016 }, { "cfg": { "lr": 0.01 }, "mean": 1.0784044712781906 } ], "full": { "mean": 0.4031560346484184, "std": 0.052733373521801685, "per_seed": [ 0.3982558250427246, 0.32343214750289917, 0.3480452001094818, 0.42022955417633057, 0.4928964078426361, 0.4031330645084381, 0.4640232026576996, 0.37523287534713745 ], "n": 8 } }, "idea": { "cfg": { "lr": 0.001 }, "mean": 0.3726405091583729, "std": 0.06837978903352387, "per_seed": [ 0.3478388786315918, 0.28696209192276, 0.32253625988960266, 0.41124576330184937, 0.4993930757045746, 0.3521648645401001, 0.44697806239128113, 0.3140050768852234 ], "n": 8 }, "comparison": { "delta_mean": -0.030515525490045547, "idea_wins": 7, "n_pairs": 8, "per_seed_diffs": [ -0.05041694641113281, -0.03647005558013916, -0.02550894021987915, -0.008983790874481201, 0.0064966678619384766, -0.05096819996833801, -0.017045140266418457, -0.06122779846191406 ], "p_value": 0.01475, "mde": 0.019566492866314403, "mde_rel_pct": 4.853330022302114, "verdict": "idea better (significant)", "system_worked": true }, "mechanism_signature": { "prediction": "gated local-global graph attention uses bounded temporal neighbors plus all-token linear statistics", "observed_mean_gate": 0.5655390620231628, "observed_local_fraction": 0.4309374988079071, "predicted_local_edge_degree": 3, "predicted_global_path": true, "confirmed": true }, "idea_sweep": [ { "cfg": { "lr": 0.001 }, "mean": 0.3726405091583729, "std": 0.06837978903352387 }, { "cfg": { "lr": 0.003 }, "mean": 0.4466487616300583, "std": 0.09400408679279276 }, { "cfg": { "lr": 0.01 }, "mean": 1.0904799774289131, "std": 0.2792352404713769 } ], "track_justification": "Sequence forecasting is structurally matched because the intervention replaces multi-token self-attention." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_bench.py", "files": [ "stage2_bench.py", "bench_report.json" ], "limitations": "Only the registered built-in sequence forecasting track was tested. Wall-clock latency, peak memory, general graph topologies, and larger datasets were not evaluated; the local branch used a fixed three-neighbor temporal pattern.", "system_verdict": "worked", "practical_verdict": "helps", "mechanism_ok": 1, "system_judged": true }