Weakest-Direction Information Margin for Latent-State Training / report_bench_2026-09-02T001424.md
Mechanism confirmed, baseline not beaten
Стенд-проверка (stage-2) · промт оператора:
(универсальный)
Ответ агента:
{ "worked": false, "confidence": 9, "verdict": "Implemented the weakest-direction curvature-margin regularizer on the structurally matched dynamics track using identical rnn_small systems and a shared learning-rate sweep. The trained-model signature was measured directly: minimum curvature and perturbation sensitivity were essentially unchanged. The idea had paired delta -2.57e-08 with permutation p=0.91285, so there was no significant benchmark win.", "metrics": { "baseline": "MSE mean 0.0122297231, std 0.0049134211, best lr=0.006", "idea": "MSE mean 0.0122296974, std 0.0049134829, best lr=0.006, paired delta=-2.57e-08, p=0.91285", "mechanism_signature": "baseline min curvature 0.0499999635, idea 0.0499999709; predicted inverse-margin ratio 0.9999998, observed perturbation ratio 0.998916; confirmed=true" }, "bench_report": { "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.006 }, "sweep": [ { "cfg": { "lr": 0.001 }, "mean": 0.1492516194 }, { "cfg": { "lr": 0.003 }, "mean": 0.03069703 }, { "cfg": { "lr": 0.006 }, "mean": 0.0124126468 } ], "full": { "mean": 0.0122297231, "std": 0.0049134211, "per_seed": [ 0.0141053172, 0.0131594678, 0.0073661068, 0.0150196953, 0.0033576291, 0.0126793338, 0.0210158844, 0.0111343507 ], "n": 8 } }, "idea": { "mean": 0.0122296974, "std": 0.0049134829, "per_seed": [ 0.0141055714, 0.0131597547, 0.0073660286, 0.0150185013, 0.0033577736, 0.0126797874, 0.0210164543, 0.0111337081 ], "n": 8 }, "comparison": { "delta_mean": -2.5698682e-08, "idea_wins": 3, "n_pairs": 8, "per_seed_diffs": [ 2.5425106e-07, 2.8684735e-07, -7.8231096e-08, -1.1939555e-06, 1.4458783e-07, 4.5355409e-07, 5.6996942e-07, -6.4261258e-07 ], "p_value": 0.91285, "mde": 5.0345351e-07, "mde_rel_pct": 0.00411664, "verdict": "no measurable effect", "system_worked": false }, "mechanism_signature": { "baseline_min_curvature": 0.0499999635, "idea_min_curvature": 0.0499999709, "baseline_trace_curvature": 1.3966932297, "idea_trace_curvature": 1.3966948986, "baseline_observed_perturbation": 7.737995e-05, "idea_observed_perturbation": 7.729608e-05, "predicted_inverse_margin_ratio": 0.9999998, "observed_perturbation_ratio": 0.998916, "confirmed": true } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_experiment.py", "files": [ "bench_experiment.py", "bench_report.json" ], "limitations": "Only the built-in dynamics track was tested; no tabular, vision, sequence, transition-mismatch sweep, larger latent-state model, or longer training schedule was evaluated. The curvature computation used a small batch subset and the observed signature had nearly identical margins, limiting evidence for a meaningful robustness effect.", "system_verdict": "partial", "practical_verdict": "no_effect", "mechanism_ok": 1, "system_judged": true }