Value-Gradient Trajectory Collocation / report_bench_2026-09-01T131251.md
Mechanism confirmed, baseline not beaten
Стенд-проверка (stage-2) · промт оператора:
(универсальный)
Ответ агента:
{ "worked": false, "confidence": 9, "verdict": "Implemented Value-Gradient Trajectory Collocation on the structurally matched dynamics track using the shared rnn_small architecture, paired seeds, equal epochs/batch/optimizer, and a parity-compliant learning-rate sweep. The mechanism signature measured on trained models confirmed 1.0 sign agreement for the feedback control prediction, but test MSE was slightly worse for the idea (0.79414 vs 0.78750), with delta_mean=+0.00664 and permutation p=0.4855. Therefore the mechanism signal is present, but the promised benchmark improvement was not observed.", "metrics": { "baseline": "dynamics/rnn_small: best lr=0.001; 8-seed test MSE mean=0.7875012234, std=0.0678897644; paired baseline values=[0.8156408072,0.7547019720,0.8308971524,0.9123226404,0.7518438101,0.7541651726,0.8133183718,0.6671198606]. Baseline sweep means: lr=0.001 -> 0.8283906430, lr=0.003 -> 0.8293724805, lr=0.01 -> 0.8328554034.", "idea": "feedback trajectory sampler with alpha=0.7, sigma=0.035, dt=0.05, lr=0.001; 8-seed test MSE mean=0.7941433117, std=0.0710090306; paired idea values=[0.8096513748,0.8031849861,0.8362289667,0.9341802597,0.7649334669,0.7608115077,0.7781288624,0.6660270691]. Idea sweep means: lr=0.001 -> 0.7941433117, lr=0.003 -> 0.8040007651, lr=0.01 -> 0.8066116720. paired delta_mean=+0.0066420883, idea wins=3/8, permutation p=0.4855." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_bench.py", "files": [ "stage2_bench.py", "bench_report.json" ], "limitations": "The completed run used a reduced but complete protocol budget of n_train=256, n_test=128, and 5 epochs per run after the initial 400-sample/14-epoch version exceeded the time limit. Only one alpha and sigma setting was tested; no broader sampler hyperparameter sweep, longer training, or comparison against residual-based adaptive sampling was run. PyTorch emitted the benchmark's existing MSE broadcasting warning because targets are shape [N] while rnn_small outputs [N,1]; this shape convention was shared by both systems.", "bench_report": { "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.001, "alpha": 0.7, "sigma": 0.035 }, "sweep": [ { "cfg": { "lr": 0.001, "alpha": 0.7, "sigma": 0.035 }, "mean": 0.8283906430006027 }, { "cfg": { "lr": 0.003, "alpha": 0.7, "sigma": 0.035 }, "mean": 0.8293724805116653 }, { "cfg": { "lr": 0.01, "alpha": 0.7, "sigma": 0.035 }, "mean": 0.8328554034233093 } ], "full": { "mean": 0.787501223385334, "std": 0.06788976443931306, "per_seed": [ 0.8156408071517944, 0.7547019720077515, 0.8308971524238586, 0.9123226404190063, 0.7518438100814819, 0.7541651725769043, 0.8133183717727661, 0.6671198606491089 ], "n": 8 } }, "idea": { "mean": 0.7941433116793633, "std": 0.07100903064007838, "per_seed": [ 0.8096513748168945, 0.803184986114502, 0.8362289667129517, 0.9341802597045898, 0.7649334669113159, 0.7608115077018738, 0.7781288623809814, 0.6660270690917969 ], "n": 8 }, "comparison": { "delta_mean": 0.006642088294029236, "idea_wins": 3, "n_pairs": 8, "per_seed_diffs": [ -0.005989432334899902, 0.04848301410675049, 0.005331814289093018, 0.021857619285583496, 0.013089656829833984, 0.006646335124969482, -0.03518950939178467, -0.0010927915573120117 ], "p_value": 0.4855, "mde": 0.0199747259489214, "mde_rel_pct": 2.5364691959529213, "verdict": "no measurable effect", "system_worked": false }, "mechanism_signature": { "idea_config": { "lr": 0.001, "alpha": 0.7, "sigma": 0.035 }, "sampler": "feedback pendulum rollout + Gaussian exploration + uniform reservoir", "prediction": "gradient sign predicts which bounded control increases model-predicted next angle", "observed_sign_agreement": 1.0, "predicted_threshold": 0.75, "confirmed": true } }, "system_verdict": "partial", "practical_verdict": "no_effect", "mechanism_ok": 1, "system_judged": true }