A reproducible Lottery Ticket Hypothesis retest on handwritten digits. The dense network is trained, globally pruned by weight magnitude to 20% of its connections, rewound to its exact initial weights, and retrained. An equally sparse random mask is trained from the same initialization as the causal control.

The saved report measures clean accuracy and Gaussian-noise accuracy for all three models. The Space runs them side by side on identical held-out examples.

The dense 6,570-parameter network reached 97.04% clean accuracy. Its 1,399-active-parameter magnitude ticket matched 97.04%, but an equally sparse random-mask control reached 97.78% and also led under Gaussian noise (93.33% versus 91.85%). On this benchmark, sparsity survived rewinding, but magnitude selection did not establish a special winning-ticket advantage.

Results are verified local artifacts; Hub deployment remains a separate step.