Why a text answer has no quorum

The contract decides whether operators agree by hashing their answers and counting equal hashes. That is the whole adjudication: no judge, no similarity score, no second model asked to compare. The first digest to reach a majority of the quorum wins, and every operator who submitted that digest is paid.

This works because three of the four answer shapes are fixed-width bytes. Two operators who both think the answer is “true” produce the byte 0x01; two operators who both score the essay 84 produce the same thirty-two bytes. There is nothing to interpret.

Bool    1 byte    0x00 | 0x01
Uint    32 bytes  abi.encode(uint256)
Choice  1 byte    the index of the option
Text    1–2048    UTF-8, whatever the model wrote

Why prose breaks it

Ask three good models the capital of Finland and you get “Helsinki”, “Helsinki.” and “The capital of Finland is Helsinki.” Three right answers, three different digests, no majority, and the contract would refund the asker and pay nobody — for a question every operator got right.

The tempting fixes are all worse than the problem. Normalising the text (lowercase, strip punctuation) catches the easy cases and fails the moment a model adds a clause. Asking a fourth model to judge whether three answers “mean the same” puts a model in charge of paying models. Embedding similarity needs a threshold nobody can defend on chain. Each of these turns a contract that counts into a contract that opines, and the whole point of the oracle is that the contract only counts.

So text is one operator

A Text request has a quorum of one, and the contract refuses any other value. The first bonded operator to reply settles it and is paid. That is a weaker guarantee than the typed shapes offer, and the site says so wherever Text appears: do not settle money on a Text answer.

What Text is for is everything a contract wants to record rather than act on: a summary posted alongside a decision, a rationale, a label for a human to read later. For those the single reply, the named model and the receipt are enough, because a reader who disagrees can see exactly who said it and run the question again.

Ask typed questions

The practical lesson is to turn the question you have into a shape that counts. “Did the deliverable meet the brief?” is a Bool. “How well, from 0 to 100?” is a Uint. “Which of these four bids fits best?” is a Choice with the options numbered in the prompt. Each of those can carry a quorum of three, and a quorum of three is the difference between “a model said so” and “three independently paid runs agreed”.

The contract’s rule is one line: count * 2 > quorum settles. Everything in this note is a consequence of keeping it that short.