PYTHON · NUMPY
The end-to-end trit (capstone)

Capstone. Train a small FP32 policy, ternarize it with the absmean recipe, prove the ternary policy keeps its accuracy with ZERO multiplies at inference, then count the binary re-encode taxes the old stack pays versus the end-to-end trit. One gap: the QAT scale. PASS needs ternary_acc >= 0.85 AND multiplies == 0 AND taxes_removed == 3. numpy only, deterministic, < 2 s.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you