Why free text breaks pipelines

The answer is right and the program is broken

Ask a model to classify something and it will happily reply "Based on the review, the sentiment appears to be positive." A person reads that and nods. label == "positive" is False, and the row goes into your database as unclassified.

The failure is not the model's. It is a missing contract: nobody said what the answer had to look like, so it looked like English.

Two halves fix it, and both are needed:

1. Ask for the shape. "Reply with exactly one word: positive, negative or neutral."
2. Check what came back. Never assume the instruction was followed — a model under a long prompt drifts, and a drifted answer that goes unchecked is a silent bad row.

ALLOWED = {"positive", "negative", "neutral"}

label = reply.strip().lower()
if label not in ALLOWED:
    raise ValueError("unexpected label: " + repr(reply))

Raising here feels harsh and is the point. A pipeline that stores "Based on the review..." in a label column fails silently and is discovered by a chart that looks odd three weeks later.

Your turn: write strict_label(reply) that normalises a reply and returns the label, or raises ValueError when it is not one of the three.

Asking for JSON

One word is the simplest contract. The next one up is JSON, and the instruction has to say so plainly — the word "JSON" in the prompt is what switches most models into that mode.

RULE = "Reply with JSON only. No preamble, no code fence."

Then parse it. json.loads raises ValueError (JSONDecodeError is a subclass) on anything that is not JSON, and that exception is the most useful signal in the whole pipeline: it fires at the boundary, on the request that broke, rather than three functions downstream where a KeyError tells you nothing about why.

Your turn: write ask_json(question) that asks for JSON and returns the parsed dictionary. Let a parse failure raise — repairing it is the next lesson, and code that swallows the error now will make that lesson impossible.