Scale Boundary

Why Exact by Hand Stops

The boundary is scale and enumeration. Real model verification needs reproducible execution, not a claim that rounded displays can be recomputed by hand.

highlighted = computed this step

Exact arithmetic still applies

A real model still uses arithmetic step by step. Counts, shapes, and discrete choices can be checked exactly when the inputs are shown.

same kind of arithmetic, many more entries\text{same kind of arithmetic, many more entries}
Scale boundaryExact counts and hand-enumeration limit.toy MLP parameter countlayerd_ind_outweightsbiastotalhidden22426output21213total params=9finite square-layer grid, not a smooth curvedd*(d+1)26420872162721001010010001001000discrete exact rows onlyreal-size integer countscomponentleftrightbiascountembedding table50000768038400000dense layer768768768590592counts exact; trained values not enumerated by handexact mechanisms on toys; scale blocks hand enumeration of real trained values; finite grid only;NOT learning; NOT generalization

The boundary is enumeration

The issue is not exactness in principle. The issue is the number of parameters and operations: too many trained values to list, inspect, and recompute by hand.

scale blocks hand enumeration\text{scale blocks hand enumeration}
Scale boundaryExact counts and hand-enumeration limit.toy MLP parameter countlayerd_ind_outweightsbiastotalhidden22426output21213total params=9finite square-layer grid, not a smooth curvedd*(d+1)26420872162721001010010001001000discrete exact rows onlyreal-size integer countscomponentleftrightbiascountembedding table50000768038400000dense layer768768768590592counts exact; trained values not enumerated by handexact mechanisms on toys; scale blocks hand enumeration of real trained values; finite grid only;NOT learning; NOT generalization

Summary

That is why the practical-phase book used a captured real trace and hash locks. For real models, verification moves from hand recomputation to reproducible execution.

hand recompute→re-execute and hash\text{hand recompute}\to\text{re-execute and hash}
Scale boundaryExact counts and hand-enumeration limit.toy MLP parameter countlayerd_ind_outweightsbiastotalhidden22426output21213total params=9finite square-layer grid, not a smooth curvedd*(d+1)26420872162721001010010001001000discrete exact rows onlyreal-size integer countscomponentleftrightbiascountembedding table50000768038400000dense layer768768768590592counts exact; trained values not enumerated by handexact mechanisms on toys; scale blocks hand enumeration of real trained values; finite grid only;NOT learning; NOT generalization