Jacobian-Frozen Stable Rollouts / report_bench_2026-09-04T152155.md

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP

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

(универсальный)

Ответ агента:

{ "worked": false, "confidence": 10, "verdict": "The full stage-2 bench was run on the structurally matched registered dynamics/rnn_small track with a tuned baseline sweep and 8 paired seeds. The Jacobian penalty reduced observed local Jacobian gain from 0.3045 to 0.2943, confirming the mechanism signal, but increased test MSE from 0.00054517 to 0.00058295; the paired permutation test gave p=0.0081 and verdict "idea worse (significant)". Therefore the idea does not transfer as a task-metric win on this benchmark.", "metrics": { "baseline": "Mean test MSE 0.00054517226817552; full per-seed values [0.0004930912400595844, 0.0004765945195686072, 0.0004673648509196937, 0.0010376222198829055, 0.000511663849465549, 0.00027233464061282575, 0.0005571767687797546, 0.0005455300561152399].", "idea": "Mean test MSE 0.0005829526708112098; full per-seed values [0.000502792012412101, 0.0004844724608119577, 0.0005921015981584787, 0.0010409810347482562, 0.000540185603313148, 0.00027245740056969225, 0.000595726422034204, 0.0006349048344418406]. Paired delta +0.000037780402635689825, p=0.0081." }, "bench_report": { "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.01 }, "sweep": [ { "cfg": { "lr": 0.001 }, "mean": 0.001864717691205442 }, { "cfg": { "lr": 0.003 }, "mean": 0.0010901302302954718 }, { "cfg": { "lr": 0.01 }, "mean": 0.0006186682076076977 } ], "full": { "mean": 0.00054517226817552, "std": 0.00020361842727135583, "per_seed": [ 0.0004930912400595844, 0.0004765945195686072, 0.0004673648509196937, 0.0010376222198829055, 0.000511663849465549, 0.00027233464061282575, 0.0005571767687797546, 0.0005455300561152399 ], "n": 8 } }, "idea": { "mean": 0.0005829526708112098, "std": 0.00020220543356011604, "per_seed": [ 0.000502792012412101, 0.0004844724608119577, 0.0005921015981584787, 0.0010409810347482562, 0.000540185603313148, 0.00027245740056969225, 0.000595726422034204, 0.0006349048344418406 ], "n": 8 }, "comparison": { "delta_mean": 3.7780402635689825e-05, "idea_wins": 0, "n_pairs": 8, "per_seed_diffs": [ 9.700772352516651e-06, 7.877941243350506e-06, 0.00012473674723878503, 3.3588148653507233e-06, 2.852175384375903e-05, 1.2275995686962257e-07, 3.854965325444937e-05, 8.937477832660079e-05 ], "p_value": 0.0081, "mde": 3.818063477530496e-05, "mde_rel_pct": 7.003407363892651, "verdict": "idea worse (significant)", "system_worked": false }, "mechanism_signature": { "baseline": { "predicted_radius_target": 0.98, "observed_local_jacobian_gain_mean": 0.3045346140861511, "observed_full_input_jacobian_gain_mean": 0.45282769203186035, "observed_local_gain_median": 0.30631518363952637, "n_samples": 32, "confirmed": true }, "idea": { "predicted_radius_target": 0.98, "observed_local_jacobian_gain_mean": 0.2943377196788788, "observed_full_input_jacobian_gain_mean": 0.44500285387039185, "observed_local_gain_median": 0.29613810777664185, "n_samples": 32, "confirmed": true }, "prediction": "The trained Jacobian penalty should reduce local input sensitivity toward its target; this was observed.", "confirmed": true } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_experiment.py", "files": [ "bench_experiment.py", "bench_report.json" ], "limitations": "The benchmark's dynamics task predicts one future angle from an 8-step window, so the implementation uses a local differentiable Jacobian stability penalty rather than a full multi-step hidden-state frozen affine rollout. Explicit spectral-radius clipping, longer-horizon forecasting, hidden-state Jacobians, and inference-speed savings were not tested.", "system_verdict": "partial", "practical_verdict": "harms", "mechanism_ok": 1, "system_judged": true }