Lattice Error-Feedback Residual Blocks

Run

/home/maxwelhelp/main/bin/python3 toy_experiment.py

This writes results.json and prints the same JSON.

What is tested

The toy residual stream uses scalar increments (a scalar is one component of the claimed componentwise result), nearest-lattice rounding, and fixed lattice spacing h.

• identity_check: verifies z_D = sum(Delta_t) - c_D and checks the nearest-rounding carry bound.
• depth_sweep: uses the adversarial constant increment Delta=0.49h. Full-state write-back accumulates a 0.49h error each layer, while feedback remains bounded by h/2.
• scale_sweep: repeats the same relative increment for several h; the normalized carry remains 0.5h and absolute error scales linearly with h.
• random_comparison: compares full-state write-back, increment feedback, and stochastic absolute-state rounding on random residual proposals.

The experiment intentionally has no learned network: it isolates the quantization mechanism and its conservation law. No claim about CIFAR accuracy or wall-clock speed is made.