Equivariant Shared-Mechanism World Model / report_bench_2026-09-04T152422.md
Mechanism confirmed, baseline not beaten
Стенд-проверка (stage-2) · промт оператора:
(универсальный)
Ответ агента:
{ "worked": false, "confidence": 8, "verdict": "The idea was implemented as a shared rnn_small pendulum predictor with a C2 equivariance penalty and evaluated on the registered dynamics benchmark. The exact C2 sanity checks passed and the trained model showed low equivariance violation, but the task improvement was not significant (paired delta -0.0002274, permutation p=0.72035), so this is not a validated benchmark win.", "metrics": { "baseline": "Best lr=0.01, epochs=10; full 8-seed MSE mean 0.0028021317, std 0.0016713125.", "idea": "Best lr=0.01, lambda=0.2, epochs=10; full 8-seed MSE mean 0.0025747306, std 0.0013220860; paired delta -0.0002274011, p=0.72035.", "mechanism_signature": { "mean_prediction_abs": 0.7621783018112183, "trained_model_c2_penalty": 0.0005555228563025594, "prediction_metric_seed0": 0.0030565059278160334, "expected_c2_penalty_direction": "lower is more equivariant", "confirmed": true } }, "bench_report": { "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.01, "epochs": 10, "lam": 0.0 }, "sweep": [ { "cfg": { "lr": 0.001, "epochs": 10, "lam": 0.0 }, "mean": 0.022081697126850486 }, { "cfg": { "lr": 0.003, "epochs": 10, "lam": 0.0 }, "mean": 0.0036966767656849697 }, { "cfg": { "lr": 0.01, "epochs": 10, "lam": 0.0 }, "mean": 0.0033956777187995613 } ], "full": { "mean": 0.0028021317266393453, "std": 0.001671312456779879, "per_seed": [ 0.001397879095748067, 0.004815699998289347, 0.0011151251383125782, 0.006254006642848253, 0.001966016599908471, 0.002434998285025358, 0.0020769357215613127, 0.0023563923314213753 ], "n": 8 } }, "idea": { "mean": 0.002574730635387823, "std": 0.0013220860264022514, "per_seed": [ 0.0030565059278160334, 0.0013654178474098444, 0.003834570525214076, 0.005198074039071798, 0.0012422980507835746, 0.0024880883283913136, 0.002220690716058016, 0.0011921996483579278 ], "n": 8 }, "comparison": { "delta_mean": -0.0002274010912515223, "idea_wins": 4, "n_pairs": 8, "per_seed_diffs": [ 0.0016586268320679665, -0.0034502821508795023, 0.002719445386901498, -0.001055932603776455, -0.0007237185491248965, 5.309004336595535e-05, 0.00014375499449670315, -0.0011641926830634475 ], "p_value": 0.72035, "mde": 0.0015692337631269659, "mde_rel_pct": 56.00142734934808, "verdict": "no significant win", "system_worked": false }, "math_check": { "c2_input_composition_max_abs": 0.0, "c2_output_composition_max_abs": 0.0, "exact_transform_identity": true }, "idea_sweep": [ { "cfg": { "lr": 0.001, "epochs": 10, "lam": 0.05 }, "mean": 0.022260770318098366 }, { "cfg": { "lr": 0.001, "epochs": 10, "lam": 0.2 }, "mean": 0.022491522133350372 }, { "cfg": { "lr": 0.001, "epochs": 10, "lam": 1.0 }, "mean": 0.020668090670369565 }, { "cfg": { "lr": 0.003, "epochs": 10, "lam": 0.05 }, "mean": 0.0037502831255551428 }, { "cfg": { "lr": 0.003, "epochs": 10, "lam": 0.2 }, "mean": 0.0036739293136633933 }, { "cfg": { "lr": 0.003, "epochs": 10, "lam": 1.0 }, "mean": 0.00422825881105382 }, { "cfg": { "lr": 0.01, "epochs": 10, "lam": 0.05 }, "mean": 0.0037559740303549916 }, { "cfg": { "lr": 0.01, "epochs": 10, "lam": 0.2 }, "mean": 0.003363642084877938 }, { "cfg": { "lr": 0.01, "epochs": 10, "lam": 1.0 }, "mean": 0.0034596045152284205 } ], "structural_match": "controlled pendulum rollout has sign-equivariant local dynamics", "mechanism_signature": { "mean_prediction_abs": 0.7621783018112183, "trained_model_c2_penalty": 0.0005555228563025594, "prediction_metric_seed0": 0.0030565059278160334, "expected_c2_penalty_direction": "lower is more equivariant", "confirmed": true } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_experiment.py", "files": [ "bench_experiment.py", "bench_report.json" ], "limitations": "The initial 4000-train/1000-test, 30-epoch configuration exceeded the 40-minute limit, so the completed protocol used 400 training and 400 test examples with 10 epochs. No held-out environment, multi-step adaptation, or larger-scale graph/world-model evaluation was performed.", "system_verdict": "partial", "practical_verdict": "inconclusive", "mechanism_ok": 1, "system_judged": true }