Covariance-Adaptive Hermite Latent Bottleneck / report_bench_2026-09-04T181313.md
Mechanism confirmed, baseline not beaten
Стенд-проверка (stage-2) · промт оператора:
(универсальный)
Ответ агента:
{ "worked": false, "confidence": 9, "verdict": "Implemented and reran the covariance-adaptive Hermite latent bottleneck on the registered tabular Friedman regression track using 8 paired seeds and matched baseline/idea hyperparameter grids. The tuned baseline achieved test MSE 7.061818897724152, while the best idea achieved 178.17990684509277; paired delta was +171.11808794736862 with permutation p=0.0081, so the idea was significantly worse. The trained latent covariance defect was q=1.0 for every seed, outside the claimed q<1 regime, and every run selected degree 0.", "metrics": { "baseline": "Best lr=0.006, weight_decay=0.0001; full 8-seed mean test MSE 7.061818897724152, std 0.3679269493562873.", "idea": "Best lr=0.006, weight_decay=0.0001; full 8-seed mean test MSE 178.17990684509277, std 9.629824242821801.", "comparison": "delta_mean=171.11808794736862, idea_wins=0/8, permutation p_value=0.0081, verdict='idea worse (significant)'." }, "bench_report": { "bench_version": 1, "track": "tabular", "model": "mlp_tiny", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.006, "weight_decay": 0.0001 }, "sweep": [ { "cfg": { "lr": 0.001, "weight_decay": 0.0001 }, "mean": 16.534404277801514 }, { "cfg": { "lr": 0.003, "weight_decay": 0.0001 }, "mean": 9.554428339004517 }, { "cfg": { "lr": 0.006, "weight_decay": 0.0001 }, "mean": 7.10407030582428 } ], "full": { "mean": 7.061818897724152, "std": 0.3679269493562873, "per_seed": [ 6.67182731628418, 6.9647216796875, 7.721733093261719, 7.057999134063721, 6.471690654754639, 6.978625297546387, 7.271554470062256, 7.3563995361328125 ], "n": 8 } }, "idea": { "mean": 178.17990684509277, "std": 9.629824242821801, "per_seed": [ 190.22879028320312, 175.91212463378906, 179.41261291503906, 165.240966796875, 190.38267517089844, 169.86196899414062, 167.13531494140625, 187.26480102539062 ], "n": 8, "cfg": { "lr": 0.006, "weight_decay": 0.0001 } }, "comparison": { "delta_mean": 171.11808794736862, "idea_wins": 0, "n_pairs": 8, "per_seed_diffs": [ 183.55696296691895, 168.94740295410156, 171.69087982177734, 158.18296766281128, 183.9109845161438, 162.88334369659424, 159.863760471344, 179.9084014892578 ], "p_value": 0.0081, "mde": 8.726735703488881, "mde_rel_pct": 123.57631695003239, "verdict": "idea worse (significant)", "system_worked": false }, "mechanism_signature": { "quantity": "trained held-out latent Hermite omitted-energy proxy", "predicted_mean_tail_scale": 1.0, "observed_mean_tail": 0.3333332120035596, "mean_q": 1.0, "degrees": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "confirmed": true }, "idea_sweep": [ { "cfg": { "lr": 0.001, "weight_decay": 0.0001 }, "mean": 224.16369247436523 }, { "cfg": { "lr": 0.003, "weight_decay": 0.0001 }, "mean": 204.7395420074463 }, { "cfg": { "lr": 0.006, "weight_decay": 0.0001 }, "mean": 178.17990684509277 } ], "protocol": { "paired_seeds": [ 0, 1, 2, 3, 4, 5, 6, 7 ], "epochs": 18, "batch": 128, "baseline_grid": [ { "lr": 0.001, "weight_decay": 0.0001 }, { "lr": 0.003, "weight_decay": 0.0001 }, { "lr": 0.006, "weight_decay": 0.0001 } ], "idea_grid": [ { "lr": 0.001, "weight_decay": 0.0001 }, { "lr": 0.003, "weight_decay": 0.0001 }, { "lr": 0.006, "weight_decay": 0.0001 } ], "baseline_selection": "sweep_baseline on seeds 0..3, full reevaluation on 0..7" } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_bench.py", "files": [ "stage2_bench.py", "bench_report.json", "stage2_stdout_rerun.txt" ], "limitations": "Only the registered tabular Friedman regression track was tested; vision, sequence, and dynamics were not evaluated. The implementation used a compact degree-0/1/2 Hermite basis rather than full total-degree coefficient transmission. The trained latents consistently had q=1.0, outside the stated theorem regime. The stored signature's confirmed field is mechanically true under the script's permissive rule, but the q<1 prediction should be regarded as unconfirmed. The idea and baseline shared the encoder and task but had different latent-head input widths, which should be equalized in a follow-up.", "system_verdict": "partial", "practical_verdict": "harms", "mechanism_ok": 1, "system_judged": true }