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

(универсальный)

Ответ агента:

{ "worked": false, "confidence": 9, "verdict": "Implemented and ran the exact doubly stochastic low-rank attention variant on the structurally matched registered sequence track. The mechanism signature was confirmed from trained-model behavior, but test MSE was worse than baseline (0.498095 vs 0.448069; paired delta +0.050026; permutation p=0.08665), so the idea did not produce a significant win.", "metrics": { "baseline": "Mean test MSE 0.44806899 over 8 seeds; baseline sweep selected lr=0.001 from {0.001, 0.003, 0.006}.", "idea": "Rank-8 low-rank doubly stochastic attention mean test MSE 0.49809538 over 8 seeds; idea wins 2/8 paired seeds." }, "bench_report": { "bench_version": 1, "track": "sequence", "model": "transformer_tiny", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.001, "rank": 8 }, "sweep": [ { "cfg": { "lr": 0.001, "rank": 8 }, "mean": 0.4039061442017555 }, { "cfg": { "lr": 0.003, "rank": 8 }, "mean": 0.5307872667908669 }, { "cfg": { "lr": 0.006, "rank": 8 }, "mean": 0.8058461993932724 } ], "full": { "mean": 0.4480689875781536, "std": 0.08669399607049459, "per_seed": [ 0.4437996447086334, 0.3383534550666809, 0.37107908725738525, 0.4623923897743225, 0.6176453828811646, 0.3899054229259491, 0.5426207780838013, 0.41875573992729187 ], "n": 8 } }, "idea": { "mean": 0.4980953820049763, "std": 0.0717662616842552, "per_seed": [ 0.5194478631019592, 0.45840978622436523, 0.43587931990623474, 0.6198630928993225, 0.5728811025619507, 0.39179709553718567, 0.535282552242279, 0.45120224356651306 ], "n": 8 }, "comparison": { "delta_mean": 0.05002639442682266, "idea_wins": 2, "n_pairs": 8, "per_seed_diffs": [ 0.0756482183933258, 0.12005633115768433, 0.06480023264884949, 0.157470703125, -0.04476428031921387, 0.0018916726112365723, -0.007338225841522217, 0.03244650363922119 ], "p_value": 0.08665, "mde": 0.05673288695634186, "mde_rel_pct": 12.661641070717113, "verdict": "no significant win", "system_worked": false }, "mechanism_signature": { "prediction": "Trained low-rank attention retains row-stochastic factors and a shared latent column marginal while using O(nr) factor state.", "observed": { "row_u": 2.086162567138672e-07, "row_v": 2.2351741790771484e-07, "shared_col": 2.5033950805664062e-06, "factor_entries": 520, "dense_entries": 1024, "observed_storage_ratio": 1.9692307692307693 }, "confirmed": true }, "protocol_notes": { "matched_track": "sequence multi-token forecast", "shared_architecture": "same input projection, positional parameters, two blocks, feed-forward widths, normalization, and output head; only attention operator differs", "n_train": 400, "n_test": 200, "epochs": 8, "rank": 8, "learning_rate_union": [ 0.001, 0.003, 0.006 ] } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_ds_attention.py", "files": [ "bench_ds_attention.py", "bench_report.json" ], "limitations": "The experiment used the small standard sequence track budget of 400 training samples, 200 test samples, and 8 epochs rather than WikiText-103 or a full-scale Transformer. Only rank 8 was tested, projection overhead and peak GPU memory were not separately timed, and the factor-state ratio was measured at sequence length 32 rather than as an end-to-end peak-memory measurement.", "system_verdict": "partial", "practical_verdict": "inconclusive", "mechanism_ok": 1, "system_judged": true }