XOR needs a bend
You will learn
Why a network needs a hidden ReLU layer to compute XOR, and what one missing bend does.
XOR is 1 when exactly one input is 1, and a comment in gft_xornet.t27 says a single linear model cannot separate it. on_comb is a 2-layer net: two hidden sums, z0 and z1, each through relu, then a weighted sum of h0 and h1. It does not train: the tests pass hand-set weights, the analytic solution W=[[1,1],[1,1]], c=[0,-1], v=[1,-2]. The browser skips all 4 tests because its runner does not know assert_eq yet; the native t27c runs all 4, all pass, none vacuous. The recording runs t27c test-report, then a one-line sed that removes relu from h1 alone, and t27c test-report again: exactly one test fails, x00: at (0,0) z1 is -1.0, and the output becomes 2.0 instead of 0. git checkout puts the line back, and sha256sum prints the same hash as before the edit. Every byte in the recording was printed by the command; only the typing is staged.