Эксперимент: Conley-Certified Latent World Model (#1155)

{ "worked": false, "confidence": 9, "verdict": "Built a compact bistable 20-dimensional-observation world-model prototype with encoder, decoder, latent transition, residual estimation, padded latent-cell graph, terminal SCC extraction, and a sanity check for residual padding. The core inclusion check passed: maximum-residual padding contained all tested true images, while half-padding failed. In the experiment, the regularized model reduced max residual from 0.04042 to 0.02604 but worsened reconstruction RMSE from 0.02426 to 0.15046 and increased terminal SCCs from 3 to 10, so no reliable certification or attractor-recovery win was observed.", "metrics": { "baseline": "max residual 0.04042; q95 residual 0.01282; reconstruction RMSE 0.02426; terminal SCCs 3; loose certificate proxy true", "idea": "max residual 0.02604; q95 residual 0.01808; reconstruction RMSE 0.15046; terminal SCCs 10; loose certificate proxy true" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json", "README.md", "article.md" ], "limitations": "Only one small deterministic bistable system and one seed were tested. The implementation uses sampled cell images rather than rigorous interval arithmetic, does not compute true Conley indices or basin-assignment accuracy, and the proposed semiconjugacy penalty duplicates the baseline one-step prediction objective. The reported certificate proxy used cell width rather than a fully rigorous isolating-set margin." }