Question 1

How is this different from asking a chat model for JSON labels?

Accepted Answer

A chat stack prompts for JSON then parses, retries, and pays for discarded tokens. Here each option is one letter, the engine reads logits at that position, merges letter variants, and returns a distribution per question.

Question 2

Will my existing Jev client work against local serve?

Accepted Answer

Same wire shape for noul, choice, and score questions, so a Jev client works unchanged against local serve. Extensions like numeric ranges, abstain slots, layouts, and large-option expansion are ignored by plain Jev clients.

Question 3

Which layout should I pick for long versus repeated work?

Accepted Answer

State-first suits long documents, question-first caches repeated heads on short streams, header and catalog list questions before state, and auto picks based on head versus state size with the resolved choice reported.

Question 4

What injection risks remain if nothing is generated?

Accepted Answer

There is no generated text to leak prompts, smuggle tool calls, or break parsers; only server-built numbers keyed by your option names leave the engine. Persuasion toward another answer is still possible but bounded.

Question 5

What happens to a choice question with more than 26 options?

Accepted Answer

The 26 letters are the whole output surface, so past that the question expands. With expand: probes, the default, each option gets an independent yes-or-no probe over the shared prefix, batched in waves, and the yes-probabilities are normalized into the returned distribution, up to 256 options. With expand: pages, candidates are split into pages of at most 26, which decodes far fewer items but gives page-conditional probabilities and no abstention.

Question 6

What is coverage, and what else does a response tell me?

Accepted Answer

Every response carries x_snap, an account of what the engine actually did: prompt tokens against tokens decoded, cached_head_tokens, shared_prefix_tokens, cache hits and misses, waves, and decode and total milliseconds. Per answer you also get coverage, which is how much of the model's raw next-token probability landed on the option letters, alongside status, confidence and typed fields. An abstain slot shows up as status abstained.

Question 7

What does it need, and what license is it under?

Accepted Answer

MIT, on macOS, Linux and Windows. It is local and deterministic, serving the same POST /v1/systemone from snap serve, so the model runs on your machine rather than at an API. Quality depends on the open model you serve and the prompt layout you pick, and answers are shaped by layout rather than by parsing. Object and array state is rendered as TOON rather than JSON, which costs about one percent fewer prompt tokens.

Snap

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community