Majority voting

best = ai[Sentiment] 5 "classify this review: {text}"

Runs 5 independent samples and returns the majority-coerced value. A quality dial for noisy classifications, in one number.

When it helps

Classification into a small set of options, where the model is usually right and occasionally not. Voting turns "right 85% of the time" into something better without changing the prompt, because the errors are usually not the same error.

When it does not

• Generation. Five different summaries have no majority. Voting needs an

answer space small enough for agreement to mean something.

• A systematically wrong model. If the model is confidently wrong about this

input, all five samples agree and voting launders the error into apparent confidence. Voting reduces variance, not bias.

The count

n is 1 to 25. Higher n costs linearly and improves sub-linearly; 3 or 5 is usually the whole benefit.

Cost

Each sample runs its own retry loop. ai[T] 5 "..." can spend up to 5 × 4 = 20 provider calls before returning.

Votes bypass the cache - identical cached responses would defeat the point - but they do count against ECKO_AI_MAX_CALLS. This is the single easiest way to multiply a bill by five without noticing, so set a budget.

Combining

Cannot combine with -> stream (there is no partial majority) or with using tools. Works with any type, including structs and enums.

Offline

In mock mode every sample returns the same deterministic value, so the majority is that value. The wiring is exercised; the variance-reduction is not, because there is no variance.