Instrument the 3×3 plus-sign image classifier

Instrumentation should explain behavior without collecting the raw material unnecessarily. For this artifact, record enough to calculate confusion matrix counts, accuracy on clean patterns, and robustness under every single-pixel flip. Include version or configuration identifiers so two runs can be compared. A log line is useful only if it answers a debugging question; dumping entire inputs creates noise and may create a privacy problem.

The artifact's user-facing goal is specific: classify a tiny pixel grid as plus or other while keeping the pixels, score, threshold, and prediction visible. Its accepted input is exactly three rows of three binary brightness values. 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 sample emits a compact metric record that a test or dashboard could aggregate.

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

truth=['plus','plus','other','other']; pred=['plus','other','other','other']
tp=sum(a==b=='plus' for a,b in zip(truth,pred)); fn=sum(a=='plus' and b!='plus' for a,b in zip(truth,pred))
print({'tp':tp,'fn':fn,'recall':tp/(tp+fn)})

Expected output: true-positive 1, false-negative 1, recall 0.5. 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.

Print the 3×3 grid, flattened indices, five selected pixels, score, threshold, and label. If a rotated or shifted plus fails, that is an invariance limitation rather than a Python bug. Enumerate all nine single-pixel flips instead of trying one convenient corruption. Shape and binary-value errors should stop before scoring so malformed images cannot masquerade as confident predictions.

At the testing and observability 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.

Use named positive and negative grids to compute confusion counts, then report every single-pixel-flip result. Accuracy alone hides whether the toy misses pluses or calls unrelated lines pluses, so include recall and false-positive count. Keep the threshold in the report and test a rotated plus separately. This evaluates a handcrafted detector, not a camera-ready vision model.

For this testing and observability 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.