Стенд-проверка (stage-2) · промт оператора:

(универсальный)

Ответ агента:

{ "worked": false, "confidence": 9, "verdict": "Implemented the Spectral Lookahead Gate on the registered dynamics/rnn_small benchmark using bench.train_model and a trained GRUCell Jacobian/readout residual. The mechanism prediction was qualitatively confirmed: mean residual 0.39864883571863174, mean lambda 0.7705069109797478, and 100% activation. However, the gate never skipped rollout, so baseline and idea had identical test MSE (delta_mean=0, p=1.0); this is not a transfer win.", "metrics": { "baseline": "Registered dynamics/rnn_small benchmark; tuned baseline sweep over lr={0.001,0.003,0.01}, tau={0.05,0.08,0.15}, horizon={4,8}; selected lr=0.01, epochs=8, tau=0.05, horizon=4. Full 8-seed MSE mean 0.0029927210562163964, std 0.0009221810424349031.", "idea": "Spectral-gated dynamics/rnn_small; best reported 8-seed MSE mean 0.0029927210562163964, std 0.0009221810424349031. Paired delta_mean=0.0, p_value=1.0, idea_wins=0/8, observed rollout-skip rate=0%.", "baseline_sweep": [ { "lr": 0.001, "horizon": 4, "mean": 0.0069296700530685484 }, { "lr": 0.001, "horizon": 8, "mean": 0.009733024751767516 }, { "lr": 0.003, "horizon": 4, "mean": 0.005332525062840432 }, { "lr": 0.003, "horizon": 8, "mean": 0.007113165338523686 }, { "lr": 0.01, "horizon": 4, "mean": 0.002964673738460988 }, { "lr": 0.01, "horizon": 8, "mean": 0.0035731843090616167 } ] }, "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": 8, "tau": 0.05, "horizon": 4 }, "full": { "mean": 0.0029927210562163964, "std": 0.0009221810424349031, "per_seed": [ 0.003302625846117735, 0.0018407043535262346, 0.002554400824010372, 0.0041609639301896095, 0.00346640357747674, 0.002362125087529421, 0.0018692960729822516, 0.0043852487578988075 ], "n": 8 }, "sweep": { "best_cfg": { "lr": 0.01, "epochs": 8, "tau": 0.05, "horizon": 4 }, "sweep": [ { "cfg": { "lr": 0.001, "epochs": 8, "tau": 0.05, "horizon": 4 }, "mean": 0.0069296700530685484 }, { "cfg": { "lr": 0.001, "epochs": 8, "tau": 0.05, "horizon": 8 }, "mean": 0.009733024751767516 }, { "cfg": { "lr": 0.001, "epochs": 8, "tau": 0.08, "horizon": 4 }, "mean": 0.0069296700530685484 }, { "cfg": { "lr": 0.001, "epochs": 8, "tau": 0.08, "horizon": 8 }, "mean": 0.009733024751767516 }, { "cfg": { "lr": 0.001, "epochs": 8, "tau": 0.15, "horizon": 4 }, "mean": 0.0069296700530685484 }, { "cfg": { "lr": 0.001, "epochs": 8, "tau": 0.15, "horizon": 8 }, "mean": 0.009733024751767516 }, { "cfg": { "lr": 0.003, "epochs": 8, "tau": 0.05, "horizon": 4 }, "mean": 0.005332525062840432 }, { "cfg": { "lr": 0.003, "epochs": 8, "tau": 0.05, "horizon": 8 }, "mean": 0.007113165338523686 }, { "cfg": { "lr": 0.003, "epochs": 8, "tau": 0.08, "horizon": 4 }, "mean": 0.005332525062840432 }, { "cfg": { "lr": 0.003, "epochs": 8, "tau": 0.08, "horizon": 8 }, "mean": 0.007113165338523686 }, { "cfg": { "lr": 0.003, "epochs": 8, "tau": 0.15, "horizon": 4 }, "mean": 0.005332525062840432 }, { "cfg": { "lr": 0.003, "epochs": 8, "tau": 0.15, "horizon": 8 }, "mean": 0.007113165338523686 }, { "cfg": { "lr": 0.01, "epochs": 8, "tau": 0.05, "horizon": 4 }, "mean": 0.002964673738460988 }, { "cfg": { "lr": 0.01, "epochs": 8, "tau": 0.05, "horizon": 8 }, "mean": 0.0035731843090616167 }, { "cfg": { "lr": 0.01, "epochs": 8, "tau": 0.08, "horizon": 4 }, "mean": 0.002964673738460988 }, { "cfg": { "lr": 0.01, "epochs": 8, "tau": 0.08, "horizon": 8 }, "mean": 0.0035731843090616167 }, { "cfg": { "lr": 0.01, "epochs": 8, "tau": 0.15, "horizon": 4 }, "mean": 0.002964673738460988 }, { "cfg": { "lr": 0.01, "epochs": 8, "tau": 0.15, "horizon": 8 }, "mean": 0.0035731843090616167 } ] } }, "idea": { "mean": 0.0029927210562163964, "std": 0.0009221810424349031, "per_seed": [ 0.003302625846117735, 0.0018407043535262346, 0.002554400824010372, 0.0041609639301896095, 0.00346640357747674, 0.002362125087529421, 0.0018692960729822516, 0.0043852487578988075 ], "n": 8 }, "comparison": { "delta_mean": 0.0, "idea_wins": 0, "n_pairs": 8, "per_seed_diffs": [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 ], "p_value": 1.0, "mde": 0.0, "mde_rel_pct": 0.0, "verdict": "no measurable effect", "system_worked": false }, "mechanism_signature": { "trained_model": "GRUCell Jacobian and learned scalar head measured on test-time batches", "predicted": "non-eigenvector transitions should activate rollout; eigen-like transitions should skip", "observed_mean_residual": 0.39864883571863174, "observed_activation_rate": 1.0, "observed_lambda_mean": 0.7705069109797478, "confirmed": true } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_spectral_gate.py", "files": [ "bench_spectral_gate.py", "bench_report.json", "bench_run.log" ], "limitations": "The built-in dynamics task is one-step scalar pendulum regression rather than threshold-event detection, so alarm recall, false alarms, lead time, covariance-calibrated z-scores, wall-clock timing, and direct FLOP measurements were not tested. The transition is a local GRU hidden-state Jacobian rather than an explicitly learned linear state-space matrix. The learned residual was nonzero for all seeds, leaving no compute-saving regime.", "system_verdict": "partial", "practical_verdict": "no_effect", "mechanism_ok": 1, "system_judged": true }