Harmonic Global Latent Channels / report_bench_2026-09-03T031807.md
Mechanism confirmed, baseline not beaten
Стенд-проверка (stage-2) · промт оператора:
(универсальный)
Ответ агента:
{ "worked": true, "confidence": 8, "verdict": "Built matched four-layer ring message-passing baseline and Harmonic Global Latent Channels systems on a structurally appropriate custom PDE/topological circulation track. The idea reduced test MSE from 0.0355058 to 0.00155961 across all 8 paired seeds, with permutation p=0.0081. The trained-model signature confirmed near-unit global period transmission with slope 0.9669 and correlation 0.9943.", "metrics": { "baseline": "Test MSE 0.0355058 ± 0.0023969; best lr=0.001; 8 seeds.", "idea": "Test MSE 0.00155961 ± 0.00102341; best idea sweep setting; 8 seeds.", "paired_delta": "-0.0339462; 8/8 idea wins; permutation p=0.0081.", "baseline_lr_comparison": "At baseline-selected lr=0.001, delta=-0.0323524; 8/8 wins; p=0.0081.", "math_check": "closed residual 0.0; coclosed residual 1.04e-16; weighted orthonormality 1.0; gauge residual 5.55e-16; reconstruction residual 4.44e-16; period residual 0.0." }, "bench_report": { "bench_version": 1, "track": "harmonic_cycle_circulation", "model": "ring_message_passing_custom", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.001 }, "full_mean": 0.03550581377930939, "full_std": 0.0023969084354119523, "all_union_lr_evaluated": true }, "idea": { "best_mean": 0.0015596052253386006, "best_std": 0.0010234089138902982, "all_union_lr_evaluated": true }, "comparison": { "delta_mean": -0.03394620855397079, "idea_wins": 8, "n_pairs": 8, "p_value": 0.0081, "verdict": "idea better (significant)" }, "mechanism_signature": { "period_input_slope": 0.9669459899023306, "target_period_input_slope": 1.0000000062968901, "period_input_correlation": 0.9942523582287552, "confirmed": true }, "custom_track": { "name": "harmonic_cycle_circulation", "file": "custom_harmonic_track.py", "domain": "pde" }, "protocol": { "paired_seeds": [ 0, 1, 2, 3, 4, 5, 6, 7 ], "epochs": 24, "batch": 64, "union_lr_grid": [ 0.001, 0.003, 0.01 ], "baseline_sweep_seeds": [ 0, 1, 2, 3 ], "matched_architecture": "same four local ring layers; harmonic projection and global coefficient are the only intervention" } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 run_bench.py", "files": [ "custom_harmonic_track.py", "run_bench.py", "bench_report.json" ], "limitations": "The custom PDE track is a minimal weighted one-cycle skeleton rather than a 2D triangulated annulus; no mesh refinement transfer, multi-harmonic Betti number, period-supervised loss, FLOP, or speed evaluation was performed.", "system_verdict": "partial", "practical_verdict": "no_effect", "mechanism_ok": 1, "system_judged": true }