Measure whether the single-neuron logic-gate trainer works

A plausible result is not yet evidence. Evaluate with binary cross-entropy, four-example accuracy, parameter gradient magnitude, and the probability margin around 0.5. The test fixture should contain an easy positive case, an easy negative or baseline case, and the boundary case most likely to flip. Separate assertions about software contracts from claims about model quality: both matter, but they answer different questions.

The artifact's user-facing goal is specific: learn weights and a bias that reproduce the OR truth table instead of hand-tuning until one example happens to pass. Its accepted input is four binary input pairs with binary OR targets. Those statements are intentionally narrower than “build an AI system.” Narrow scope lets us inspect every input and expected result, and it prevents a toy result from being presented as a production claim. The runnable check below turns one success criterion into an assertion, so a regression exits loudly.

Save this as lesson.py and run python3 lesson.py. It uses only the language standard library, so the example is reproducible offline.

pred=[0.06,0.98,0.98,1.0]; truth=[0,1,1,1]
labels=[int(p>=.5) for p in pred]
assert labels==truth
print('accuracy',sum(a==b for a,b in zip(labels,truth))/4)

Expected output: accuracy 1.0. Exact floating-point formatting may vary slightly, but the asserted behavior must not. Read the output as evidence about this stage, not merely proof that the interpreter started.

Trace z, sigmoid probability, error, and each parameter update for one OR example. If loss stays near 0.693, confirm that the bias changes and that labels are numeric zero or one. Clamp only the exponential input for numerical stability; do not clamp the learned probability so aggressively that gradients disappear. When rounded accuracy looks perfect, inspect probability margins to catch a brittle decision near 0.5.

At the evaluation stage, save the smallest failing fixture beside the expected result. Change one cause at a time and rerun the exact command printed above; that makes the repair reviewable and keeps this chapter's progressive artifact reproducible.

Evaluate all four truth-table rows every time—there is no reason to sample a four-row domain. Show cross-entropy, exact accuracy, minimum distance from 0.5, and learned weights. Compare against constant-zero and constant-one baselines. The neuron has mastered OR only when every row passes and the probabilities are comfortably on the correct side of the threshold.

For this evaluation page, preserve the fixture and result as evidence for the next page. Label observations separately from conclusions: a passing assertion establishes the behavior it names, while broader usefulness requires the chapter's full evaluation set and stated operating limits.