Self-Consistency & Structured Output

Zero to GPT's sampling lesson covered temperature. At temperature 0 a model is deterministic, same input, same output, every time. Above 0, sampling introduces real randomness, and the same prompt can produce genuinely different reasoning paths on different runs. Self-consistency turns that randomness into a strength. Instead of trusting one sample, generate several independent ones and take the majority answer.

All five samples reasoned through the problem slightly differently, different phrasing, different order of stating intermediate numbers, but converged on the same final answer, 12. That convergence is itself informative, when independent reasoning paths agree, you have real evidence the answer is right, not just a single confident-sounding guess. If two or three of the five had landed on a different number, majority vote would have caught the outlier, a wrong reasoning chain that, sampled once, would have looked exactly as confident as a right one. Self-consistency's whole value proposition is trading extra inference calls for exactly that safety net.

A different, more mundane reliability problem is getting a model's output into an exact shape your code can parse. This is the same lesson as zero-shot vs. few-shot from lesson 2, applied to a schema instead of a label.

Without a schema, the model invented its own reasonable-but-unpredictable JSON shape, a pros array that wasn't asked for, a max_rating field split out from rating, and the whole thing wrapped in a markdown code fence. With an exact schema spelled out in the prompt (field names, types, and "reply with only the JSON"), the output matched precisely, no code fence, no extra fields, ready to parse. If your code depends on a specific shape, specifying that shape in the prompt is far more reliable than hoping the model infers it.