Jev and System One models

Select, don't generate

Most decisions in a retrieval pipeline are choices from a closed set. Is this the contents page? Which page does this entry start on? Does this section answer the question? Which sentence on this page supports this claim? Each can be answered with a probability or a pick from a list.

Vectorless sends every such decision to a System One model and keeps a generative model for the one thing that needs writing: the final answer.

If a decision can be written as a closed set of options, it goes to the judge.

The argument for this, and what it gives up, is in the engine's Judgement, not generation note (2026-09-25). In short: a chat model asked for a decision produces a plausible-sounding decision shaped by its own sampling, and an embedding decides before the question exists. A System One model reads the question and the candidate together and returns the decision directly. What is given up is deliberation: Jev decides one step at a time and cannot plan across sections. Multi-hop questions have not been measured.

What Jev is

Jev is TypeSafe's System One model, a hosted API.


Model alias	`jev-latest`
Endpoint	`api.typesafe.ai/v1/systemone`
Output	Calibrated probabilities and choices; no prose
Price	$0.042 per million input tokens; output is free
Size	Not published

One call carries a shared state and N typed questions about it:

Question type	Returns
Noul	The probability of yes
Choice	One option, plus the probability distribution over all options
Score	A position on an ordered scale, plus the distribution

The questions in one call are answered independently. In probes on 2026-09-25, latency barely changed with the number of questions in a request, so the engine packs many questions into each call.

What Jev decides in Vectorless

Stage	Decision	Type
Ingest	Which page is the contents page	Noul per page
Ingest	Which candidate lines are real contents entries	Noul per entry
Ingest	Which page each entry starts on	Noul per candidate page code found; the most probable wins
Ingest	Which sub-headings to keep when splitting a long section	Noul per heading
Query	Which sections hold the answer	Noul per section
Query	Which pages hold the answer (skim, then full read)	Noul per page
Answer	Which sentence or table row supports each claim	Choice, with a `none` option

Code lists the candidates in every case. Jev never writes a page number, a title or a quote; it picks one that code found in the document.

How it is called

The engine talks to Jev through llmgate, which has a Judge interface beside its chat Client interface. A chat-completion API has roles, messages and an assistant turn; a System One model has none of those, so it gets its own interface.

Known llmgate issue

On llmgate v0.5.0 the TypeSafe client's context guard rebuilt its tokenizer on every count (about 256 ms each). That made a query take a median 50.5 s, almost all of it our own CPU; Jev itself answered in about 0.6 s. The fix is in llmgate PR #19, which was still open on 2026-10-11. The engine pins that branch. See Benchmarks.

What needs a TypeSafe key

Set TYPESAFE_API_KEY (or VLE_TYPESAFE_API_KEY / VLS_TYPESAFE_API_KEY) to turn Jev on. TYPESAFE_BASE_URL and TYPESAFE_MODEL override the endpoint and the model.

Feature	With a key	Without a key
Contents tree at ingest	Jev decides; no generative call	Your generative model decides, one call per page; minutes slower on long filings
Leaf splitting	On (sections over 20 pages)	Off
`judgewalk` navigation	Jev ranks sections and pages	Replaced by `treewalk`: your chat model walks the pages in up to eight generative hops
Pincite quote choice	Jev picks among the shortlist	Picked by word overlap
Answer text	Your generative model	Your generative model

If Jev refuses a request (no credit, a rejected key, an outage), the server binary serves that query with treewalk instead and logs retrieval: primary model unavailable, served by fallback. This was added after the TypeSafe account ran out of credit on 2026-10-08 and every production query failed for as long as it lasted.

Your documents leave your network

With a TypeSafe key set, document text is sent to TypeSafe's API. "Nothing leaves your network" is not true while Jev is on. Your generative model provider also receives the evidence pages when it writes the answer.

The figures on Benchmarks were measured with Jev on. The only fallback that was measured is treewalk, as the same-model control in the PageIndex comparison.