Answer with one of five words

01 · Role in the system

What Jev does here

One request to Jev (through Vercel AI Gateway) evaluates three questions in parallel: kelime, a Choice among the five phrases - aynen, eyvallah, ne alaka, sıkıntı yok, yani; enerji, a 0-2 score; and soruSayilir, the probability that the reply reads as a question. The word is the reply itself, and code derives the punctuation from the two numbers: high question probability yields ? ?? ?! ...?, high energy yields ! !! !!!, low yields . .. or nothing. The five-word constraint is not a prompt request but the type system: the answer is typed as a five-member union in TypeScript, and adding a word to the constant widens the API option and the type together. The README shows the raw response with per-word probabilities and per-question confidence.

A vocabulary closed by the type system

The toy answers everything you type with one of five fixed Turkish phrases, and the constraint is not a prompt request — the answer is typed as a five-member union in TypeScript, so adding a word to the constant widens the API option and the type together. One gateway call evaluates three questions in parallel: which phrase, an energy score from zero to two, and the probability that the reply reads as a question.

Punctuation is then derived in code from the two numbers — high question probability yields question marks and interrobangs, high energy yields exclamation tiers, low yields a period or nothing. The model chooses words; grammar stays a deterministic function of judgments.

What a five-word toy actually demonstrates

The record says plainly that the interesting part is the closed typed answer space, not the conversation, and the limitations note the gateway routing rather than the native endpoint plus a TLS handshake flake observed during verification before the demo succeeded.

As research-cluster evidence it is the smallest possible illustration of a serious idea: if the answer space is a union type, hallucination outside it is unrepresentable. Scaled up — as the word-tree record in this cluster attempts — the same idea meets its limits, which is precisely what makes the pair of records worth reading together.

02 · Control boundary

Where Jev sits

One experimental_evaluate call with three parallel questions (Choice, score, boolean); code owns punctuation derivation and the typed union keeps the vocabulary closed.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• A toy with a five-word vocabulary; the interesting part is the closed typed answer space, not the conversation.
• Calls go through Vercel AI Gateway (experimental_evaluate), not the native TypeSafe endpoint.
• The interface and README are Turkish; the demo occasionally failed TLS handshakes during verification before succeeding.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• mahmut-gundogdu ↗Community · github · public · checked 2026-09-22
• mahmut-gundogdu ↗Community · demo · public · checked 2026-09-22