Эксперимент: Faithful Latent Fixed-Point Solver (#428)

{ "worked": true, "confidence": 9, "verdict": "Built a faithful latent fixed-point solver with trainable linear E, D, and T, explicit encoder/decoder intertwining losses, latent rollouts, and decoded fixed-point validation. The mechanism manifested: the exact contraction boundary matched the prediction at gamma=1, the rank-2 manifold had near-zero reconstruction residual only for m>=2, and observed rollout rates matched gamma exactly. The learned m=2 model achieved approximately 5e-8 faithfulness and fixed-point residuals, while m=1 retained substantial residual; the latent solver matched the 12-step baseline on this easy toy but did not improve its output error.", "metrics": { "baseline": "12 S evaluations; fixed-point RMSE 0.04168; output update residual 0.0075086", "idea": "0 S evaluations and 12 T evaluations; fixed-point RMSE 0.04172; output update residual 0.0075163; learned m=2 faithfulness residuals approximately 5e-8; decoded rate 0.8026 vs predicted 0.82" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "Only a linear 10D synthetic system with an exactly rank-2 invariant manifold was tested. No nonlinear DEQ, Anderson acceleration, wall-clock profiling, noisy or perturbed initial-state divergence study, or larger-scale memory/FLOP analysis was performed; the learned decoded rate was slightly below the dominant gamma because the secondary mode dominates late trajectory differences." }