The Decision Layer Doesn't Replace the Governance Layer: What Jev Means for Legal AI.

Part 1 covered the headline: a new kind of model, TypeSafe AI's Jev, has spent the past ten days going viral in AI circles for reasons that have nothing to do with writing text — it doesn't draft, summarize, or converse, it answers structured questions, fast and cheap, for exactly the kind of repeated judgment call an eDiscovery team runs by the thousands. Here's the fuller picture: what Jev actually is, why it's a real primitive for legal work, why it's not ready to be trusted blindly yet, and why none of that changes what your firm's AI governance still has to cover.

It's also worth being precise about what Jev actually is, what it isn't, and what it changes — and doesn't — for the teams responsible for governing how their firm's AI touches privileged and client-confidential material.

What Jev actually does

Jev is what TypeSafe AI calls a “System 1” model, borrowing the fast/slow framing from Kahneman: instead of generating text token by token the way an LLM does, it takes a document (the “state”) and a set of pre-defined questions with fixed answer choices, and returns a decision for each one — a choice, a score, or what TypeSafe AI is calling a “noul”: a calibrated probability between 0 and 1 for a yes/no question. Calibrated is the operative word. If Jev returns .80 across a batch of documents, the claim is that roughly 80% of them are actually responsive — a number a review team could use as an automatic-code threshold rather than a rough guess.

Because it scores a fixed set of response options in parallel instead of generating free-form output, early adopters are reporting it running dozens of classifications per document in real time, at a fraction of the cost of running the same set of questions through an LLM. Nathan Reff and Benjamin Sexton's team at JND Legal Administration, who got early access, put real numbers on it: dramatically cheaper, faster, and — notably for anyone who has watched an LLM change its mind between two runs on the exact same document — more consistent.

None of that makes it a general-purpose tool. TypeSafe AI is upfront that Jev reads questions literally, can't do multi-step reasoning, and isn't suited to counting, arithmetic, date comparisons, or generating text of any kind. And the team that's actually tested it on real legal tasks is careful not to oversell it: cheaper, faster, and more consistent are not the same claim as better, and JND's own writeup stops short of calling it production-ready — that's benchmarking work still in progress.

Why eDiscovery teams should still pay attention

The use case writes itself once you separate “decide” from “generate.” Responsiveness review, privilege screening, classifying a document set against every request in an RFP — these are, at their core, thousands of repeated structured judgments, which is exactly the shape of task Jev is built for. If the cost and speed claims hold up at scale, it's not hard to see how “reasonable” discovery burden arguments shift, or how keyword search gets displaced as the default first pass on a large corpus.

It's also a preview of a broader pattern worth watching: cheap, narrow decision models sitting upstream or downstream of an LLM in the same pipeline — an LLM scoping which questions actually need asking, a decision model running those questions across the full population, a human reviewer picking up only what falls below a confidence threshold. That's a genuinely different shape of legal AI workflow than “one chatbot handles everything,” and it's arriving at the same moment the pricing floor for legal AI generally is dropping — smaller firms that were priced out of seat-minimum enterprise tools are exactly the teams a $0.042-per-million-token classifier could open real capability to.

What doesn't change

Here's the part that's easy to lose in a viral launch: none of this touches the question of where the document actually goes.

Whether it's an LLM drafting a brief or a decision model scoring a document against a discovery request, the document still has to leave wherever it currently lives and land somewhere a model can read it. For a law firm, that document is very often privileged, client-confidential, or subject to a specific data-handling obligation the firm made to that client. A faster, cheaper decision layer doesn't relax any of that — if anything, cheaper classification means more documents flowing through more models more often, which makes the question of where they're allowed to go more urgent, not less.

That's the layer we build at Cortega, and it's model-agnostic by design — we don't care whether what's on the other end of a request is a frontier LLM, a research agent, or a new category of model like Jev that didn't exist two weeks ago. Cortega sits in the path between a firm's applications and whatever's processing their documents, and routes each request under the firm's own data policy: privileged and proprietary work stays on local or on-premises models, and only what's approved by policy goes to an outside provider. Every request that crosses that boundary is logged, so when a general counsel or a client asks what touched a document and where it went, there's an answer grounded in an audit trail rather than a best guess.

The same posture applies regardless of which kind of model shows up next. A new architecture that's faster or cheaper than what came before is good news for legal ops — it's also one more thing a firm's governance layer needs to already cover, rather than something to bolt on after the fact once it's already live inside a matter.

The honest read

Jev is genuinely interesting, and the early numbers from teams who've actually run it against real legal tasks are worth watching closely. It's also barely out the door — TypeSafe AI announced it on September 15 — its internal architecture is undisclosed, its own vendor reports 68% accuracy on its own evaluation, and competing open-source versions are already showing up. That's not a knock — it's just where a model this new actually is. The right move for most legal teams right now is the same one JND is taking publicly: test it, benchmark it against real matters, and don't take “fast and cheap” as a substitute for “validated.”

Whatever comes out of that testing, one thing won't change: a firm's AI governance can't be scoped to “the chatbots we know about today.” It has to hold for whatever shows up next — a new LLM, a new agent framework, or a new kind of model built to do one thing very fast. That's the layer worth getting right before the next Jev shows up.

Sources: Michallynn Demiter, Nathan Reff, and Benjamin Sexton, “Meet the ‘Decision Models’: What Jev Might Mean for Legal Services” (LinkedIn, Sept 23, 2026); AI Supremacy, “The Most Viral New AI Model Is Not an LLM: Jev” (Sept 2026); Alexis Keenan, LinkedIn post on legal AI pricing (Sept 2026).