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

(универсальный)

Ответ агента:

{ "worked": false, "confidence": 9, "verdict": "Excitation-gated training was evaluated on the structurally matched built-in dynamics track with the shared rnn_small architecture and standard test MSE. The idea slightly reduced mean MSE, but the paired improvement was not significant (delta_mean=-7.147682481445372e-07, p=0.93915), so there is no demonstrated benchmark win. The mechanism proxy was numerically consistent, but it is only an empirical-Fisher inverse proxy rather than direct learned calibration-variance evidence.", "metrics": { "baseline": "Best fixed-decay configuration lr=0.01, decay=20.0; full 8-seed mean test MSE=0.0014148428163025528.", "idea": "Best gated configuration lr=0.01, decay=20.0; full 8-seed mean test MSE=0.0014141280480544083; paired delta=-7.147682481445372e-07; 4/8 paired wins; permutation p=0.93915; verdict=no measurable effect." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_excitation_gated.py", "files": [ "bench_excitation_gated.py", "bench_report.json", "bench_run.txt" ], "limitations": "Only the built-in dynamics track was tested. The implementation uses action excitation as a proxy rather than a learned calibration head with autodiff Jacobians, and the mechanism signature uses an empirical-Fisher inverse proxy rather than direct calibration-error variance. No sensor-frame rotation, online deployment, broad noise sweep, or alternative dynamics task was tested.", "bench_report": { "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.01, "decay": 20.0 }, "sweep": [ { "cfg": { "lr": 0.001, "decay": 3.0 }, "mean": 0.004313088022172451 }, { "cfg": { "lr": 0.001, "decay": 8.0 }, "mean": 0.004310200281906873 }, { "cfg": { "lr": 0.001, "decay": 20.0 }, "mean": 0.004303071822505444 }, { "cfg": { "lr": 0.003, "decay": 3.0 }, "mean": 0.0034836910490412265 }, { "cfg": { "lr": 0.003, "decay": 8.0 }, "mean": 0.0034992330765817314 }, { "cfg": { "lr": 0.003, "decay": 20.0 }, "mean": 0.003510375740006566 }, { "cfg": { "lr": 0.01, "decay": 3.0 }, "mean": 0.0014869928127154708 }, { "cfg": { "lr": 0.01, "decay": 8.0 }, "mean": 0.0014831181906629354 }, { "cfg": { "lr": 0.01, "decay": 20.0 }, "mean": 0.0014822576777078211 } ], "full": { "mean": 0.0014148428163025528, "std": 0.0003996130386047582, "per_seed": [ 0.0009607335086911917, 0.0018558382289484143, 0.0010648592142388225, 0.002047599758952856, 0.0008689594687893987, 0.0014588737394660711, 0.001647297409363091, 0.0014145812019705772 ], "n": 8 } }, "idea": { "per_seed": [ 0.0009361208067275584, 0.0018451964715495706, 0.0010599215747788548, 0.0020810049027204514, 0.0008966241730377078, 0.001395710976794362, 0.0016480678459629416, 0.0014503776328638196 ], "mean": 0.0014141280480544083, "details": [ { "seed": 0, "metric": 0.0009361208067275584, "cert_rate": 0.995, "mean_fisher": 2874.3630795478816, "final_fisher": 3222.6982116699214 }, { "seed": 1, "metric": 0.0018451964715495706, "cert_rate": 0.995, "mean_fisher": 2929.559712052345, "final_fisher": 2800.8539199829097 }, { "seed": 2, "metric": 0.0010599215747788548, "cert_rate": 0.995, "mean_fisher": 2862.863855719566, "final_fisher": 2856.386566162109 }, { "seed": 3, "metric": 0.0020810049027204514, "cert_rate": 0.995, "mean_fisher": 2946.144814610481, "final_fisher": 2903.7775039672847 }, { "seed": 4, "metric": 0.0008966241730377078, "cert_rate": 0.995, "mean_fisher": 2942.1488319635387, "final_fisher": 3103.5394668579097 }, { "seed": 5, "metric": 0.001395710976794362, "cert_rate": 0.995, "mean_fisher": 2841.0421735644336, "final_fisher": 2858.1001281738277 }, { "seed": 6, "metric": 0.0016480678459629416, "cert_rate": 0.995, "mean_fisher": 2934.195323467254, "final_fisher": 2973.991966247558 }, { "seed": 7, "metric": 0.0014503776328638196, "cert_rate": 0.995, "mean_fisher": 2942.573763251304, "final_fisher": 3015.8184051513667 } ], "cfg": { "lr": 0.01, "decay": 20.0 }, "method": "gated" }, "comparison": { "delta_mean": -7.147682481445372e-07, "idea_wins": 4, "n_pairs": 8, "per_seed_diffs": [ -2.46127019636333e-05, -1.0641757398843765e-05, -4.937639459967613e-06, 3.340514376759529e-05, 2.7664704248309135e-05, -6.316276267170906e-05, 7.704365998506546e-07, 3.579643089324236e-05 ], "p_value": 0.93915, "mde": 2.807584254353381e-05, "mde_rel_pct": 1.984378916161526, "verdict": "no measurable effect", "system_worked": false }, "mechanism_signature": { "prediction": "calibration variance scales as sigma^2 / Fisher", "predicted_vs_observed": [ { "seed": 0, "fisher": 2874.3630795478816, "predicted_inverse": 1.391612642279407e-05, "observed_inverse_proxy": 1.391612642279407e-05 }, { "seed": 1, "fisher": 2929.559712052345, "predicted_inverse": 1.3653928894310687e-05, "observed_inverse_proxy": 1.3653928894310687e-05 }, { "seed": 2, "fisher": 2862.863855719566, "predicted_inverse": 1.397202312645294e-05, "observed_inverse_proxy": 1.397202312645294e-05 }, { "seed": 3, "fisher": 2946.144814610481, "predicted_inverse": 1.3577065119688808e-05, "observed_inverse_proxy": 1.3577065119688808e-05 } ], "ratio_mean": 1.0, "confirmed": true, "note": "Observed quantity is the trained-model empirical Fisher proxy, not an oracle calibration error." }, "idea_sweep": [ { "cfg": { "lr": 0.01, "decay": 20.0 }, "mean": 0.0014141280480544083 }, { "cfg": { "lr": 0.001, "decay": 3.0 }, "mean": 0.005027921928558499 }, { "cfg": { "lr": 0.001, "decay": 8.0 }, "mean": 0.005027921928558499 } ], "protocol_note": "Matched dynamics track; same rnn_small, data, epochs, batch, learning-rate/decay union, and standard test MSE. Only training-time excitation differs." }, "system_verdict": "partial", "practical_verdict": "no_effect", "mechanism_ok": 1, "system_judged": true }