Build the decision layer. Own the bar and the boundary.
It is week five. The demo routes chat, code, and image work to different models. The room likes the latency. Nobody can say who owns the rule that keeps client data on the box, or which confidence number forces a safe default instead of a guess. You did not inherit a chatbot problem. You inherited an unowned decision layer.
TypeSafe positions Jev as a System One model founded by Dio Almeida (ex-OpenAI, InstructGPT): no free text; about $0.042 per million input tokens; output free. Witteveen’s useful line for operators is plain: think of it as an if-statement that understands language.
Situation. Newly appointed CEOs of US and UK mid-market firms in the £250M-£2B / $300M-$2.5B range inherit agent demos that call OpenJevs and local judges exist so the same typed contract can run on the machine when trust demands it. Approximation is not the same as TypeSafe’s calibrated model. Thresholds tuned on an open stand-in must be re-checked on the production judge. The CEO still owns the bar and the boundary.
Action. Take three moves from the walkthrough without writing the code yourself.
- Name the decision layer. Ask for one diagram of live LLM calls. Mark each step as classification, difficulty, privacy gate, or true synthesis. If the stack cannot show that split, you do not have an operating model. You have prompts.
- Put typed decisions in front of expensive calls. Require choice, score, or yes-no before any generative spend. Route by lane and difficulty. Escalate only when the work needs synthesis or hard reasoning. Use confidence to force a safe default when the judge is unsure. That is policy, not theatre.
- Own where the judge runs. Hosted Jev when calibration and speed earn the call. An OpenJev-style local judge when client or credential data must stay inside the firm. Name the owner of the confidence thresholds and the local-or-cloud rule. Review both with ops, security, and finance in the first 90 days.
Embed one principle and keep it visible: a typed answer only helps if someone named can change the threshold and the boundary.
That is the CEO read of Witteveen’s build. Not a duplicate of volume routing or client-send gates. A decision layer you install so model spend, model risk, and data locality all have an owner.
If you want that split mapped onto the stack you inherited, request a private briefing at ceo@johncorr.com or johncorr.com/briefing.
frontier models for almost every step. Classification, difficulty, and privacy checks are buried inside prose. Finance sees one opaque token bill. Security cannot point to a named boundary between local and cloud. Boards ask for AI progress. Nobody owns the front door that decides which model, if any, may run.
Complication. Reasoning models are the wrong default for high-volume verdicts. They are slow and priced for generation. Confidence, when it appears, is treated as a demo novelty rather than policy. Privacy checks that call a hosted classifier have already leaked the prompt. That is the trap Witteveen flags in the build: if you ask the cloud whether the text is private, you have already sent the private text.
Witteveen’s walkthrough is a build, not a product tour. He wires a local router: state in, typed questions out to TypeSafe AI’s Jev (or later to an open local stand-in), then branches to a small local model, a cloud reasoning model, or a local image model. Three question types do the work in one pass: choice for the lane, score for difficulty, yes-no (null) for private data. Answers come back typed, with probabilities and confidence, in the tens to hundreds of milliseconds.