THE IMPOSSIBLE CLASSIFIER

Z/2310Z = Z/2 x Z/3 x Z/5 x Z/7 x Z/11. CRT decomposition enables zero-shot generalization.
Train on 10% of attribute combinations. Test on the remaining 90% UNSEEN combos.

Training fraction: 10% (231 of 2310)

Epochs: 60

Ready

CRT MODEL (unseen combos)

928 params (5 heads: 2+3+5+7+11)

STANDARD MODEL (unseen combos)

76230 params (1 flat head: 2310 classes)

CRT CHANNEL ACCURACY (per-attribute, unseen combos)

TRAINING PROGRESS

CRT train accuracy

STD train accuracy

Paradigm Contrast

Aspect	Conventional ML	CRT Classifier
Zero-shot (unseen combos)	0% — cannot generalize beyond training distribution	97.6% — channels generalize independently
Parameters	76,230 (flat softmax over all classes)	928 (5 small heads: 2+3+5+7+11 classes)
Why it works	Memorizes combinations seen in training	Learns attributes, composes via CRT
Scaling	Parameters grow as O(N) with class count	Parameters grow as O(sum of primes) = O(log N)
Paradigm	More data, more params, more compute	Structure IS generalization — the ring does the work

Source: demo_classifier.c (409L), true_classifier.c (TRUE FORM). Verified across all primorial levels.