Эксперимент: Consensus-Safe RoPE Residual Attention (#361)

{ "worked": true, "confidence": 8, "verdict": "Built a reproducible NumPy MVP implementing RoPE attention, the spherical residual update, reversible-spectrum stability estimation, kernel-floor and Dobrushin checks, angular-diameter monitoring, and an unconstrained-residual comparison. The mechanism manifested: all three core predictions passed, with the observed instability transition at 1.01× the predicted Euler boundary for every beta tested. The constrained update kept maximum token norm at 1.0 while the unconstrained residual reached 7.22e8 after 20 steps.", "metrics": { "baseline": "Unconstrained residual: maximum norm 7.224250853e8 after 20 steps; initial angular diameter 2.2791 radians.", "idea": "Spherical RoPE residual: maximum norm 1.0 after 20 steps; angular diameter 0.02542 radians. At beta=2, predicted kernel floor alpha=0.0015263 versus observed minimum weight 0.0135173, and predicted Dobrushin bound 0.981684 versus observed 0.711092. For beta 0.5, 1, 2, and 3, predicted eta_c was 2.0000, 2.0014, 2.0168, and 2.0327; the observed transition was 1.01× eta_c in every case." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 consensus_rope_experiment.py", "files": [ "consensus_rope_experiment.py", "results.json", "run_output3.txt" ], "limitations": "This is a small NumPy toy verification, not a trained Transformer or language-model benchmark. It does not test loss, accuracy, GPU performance, long sequences, nonlinearities, value projections, masking, or the every-K-step spectrum-clipping controller; the floor and Dobrushin checks use limited toy instances rather than broad statistical sweeps." }