Эксперимент: Lattice Error-Feedback Residual Blocks (#355)

{ "worked": true, "confidence": 9, "verdict": "Built and numerically verified scalar lattice error-feedback residual updates. The conservation identity held to 4.4e-16, baseline full-state error grew with depth at 0.49h per layer as predicted, while feedback error stayed below h/2; carry magnitude scaled linearly with h and remained 0.5h normalized. The mechanism clearly manifested, although this does not establish end-to-end neural-network accuracy or speed gains.", "metrics": { "baseline": "For Delta=0.49h, depth-512 absolute error=250.88h and fitted depth slope=0.4900000000000001h/layer; random-increment mean absolute error at depth 512=2.3713 with h=0.5.", "idea": "Identity error=4.44e-16; maximum normalized carry=0.5; depth-512 absolute error=0.12 with h=1; random-increment mean absolute error at depth 512=0.1254 versus baseline 2.3713; carry/error scaled linearly with h." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 toy_experiment.py", "files": [ "toy_experiment.py", "results.json", "README.md" ], "limitations": "Only a scalar toy residual stream was tested, with fixed nearest rounding and no learned CNN, ResNet, Transformer, CIFAR data, clipping stress test, training dynamics, GPU execution, memory accounting, or wall-clock speed measurement. The bounded-carry result assumes no saturation and does not test finite-range int4/int8 state interactions." }