Jev & Clef: The Deciders

Welcome to the lesson. Today, Jev and Clef — the deciders. And this one comes with a toy. While you listen, there is a live page where you can fire real decision questions at a real model running on Cloudflare, and watch the odds come back. Clef Lab, at clef-lab dot twenty-nine court dot com.

Before the toy, give me the one-sentence version. What is a decider?

A chat model writes. A decision model picks. Jev and Clef never produce a word of prose, an essay, or an explanation. You hand one a situation, and a short list of questions with fixed answer shapes, and it hands back the pick, the odds on every option, and a confidence number. That is the entire trick.

Fixed answer shapes. How many shapes are there?

Three. Yes or no. Pick one, from a list you supply. Or score it, on a scale you define. That is all it will ever say. And honestly, that restraint is the point.

Why is that a big deal? Chat models answer questions all day.

Because of what the work actually is. Psychologists split thinking into System One — fast, intuitive, the gut call — and System Two — slow, deliberate reasoning. TypeSafe named Jev after that split. A chat model is a System Two machine. And here is the number that matters: when we audited the calls our own coordinator layer actually makes, ninety-three point eight percent of them were decisions, not writing. Route this here. Allow it or block it. Urgent or not. Which of these three.

So we have been hiring a novelist to work a toll booth.

Brilliant, slow, expensive — and prone to improvising. The doctrine writes itself. Stop paying novelist rates for toll-booth work. Let the decider take the bounded many, and save the big brain for the hard few.

Walk me through one call, concretely.

Take an email. Quote: your invoice is overdue by sixty days. Pay immediately, or we escalate to collections. That email is what we call the state — the thing being judged. Then three questions in one call. Does this email threaten the reader, yes or no. What is the tone — warm, neutral, firm, or hostile. And how polite is it, on a scale from rude, to blunt, to neutral, to polite, to warm.

And what came back?

Tone: fifty-five percent firm, thirty-six percent hostile, almost nothing warm. Threatening: about a one-in-three chance of yes. And the detail I love — the confidence was zero point two five. Low. The model was genuinely torn between firm and hostile, and it told us so.

A chat model would have just said firm, and sounded sure.

Every time. That honesty is the product. A decider shows its work as odds. And one more subtlety, because it bites people: the score is a position on your scale, not a percentage. That email scored zero point eight four on the politeness scale, which sounds lovely until you remember the scale runs rude, blunt, neutral, polite, warm. Zero point eight four means between rude and blunt. Quite rude. The placement is the meaning.

Okay. So it can be torn. What happens when it is torn and the stakes are real?

That is the craft around the call. Thresholds: in our own rule system, anything under zero point six confidence gets flagged and escalated instead of executed. Abstention is a designed outcome — none of the above, a human should look — not a failure. Fallbacks: if the decider is down or unsure, the work drops to the older, known path. And the doctrine splits by stakes. Advisory work fails open, so a missing suggestion never blocks anyone. Routing and money fail closed. No confident decision, no movement.

And the ladder.

Cheap decider first. The few percent it flags climb to a big chat model. The genuinely ambiguous reach a person. Each rung costs roughly ten times the one below, and handles a tenth of the volume.

Now the two names. Jev first.

Jev is TypeSafe's hosted decision model, the one that defined the category. Current version, one point one four. Two things to know. First, it is hosted only — you send your data to their A P I, and the weights have never been released. Second, the famous numbers — up to two hundred times faster, four hundred times cheaper than a frontier chat model — come from a launch video. A presenter's claim, not a benchmark anyone here ran. The direction is believable. The multipliers are marketing.

And the residency line.

The Principal's own ruling: anything whose data cannot leave the building cannot use a hosted-only model. So Jev is a fine tool for public or disposable data, and a non-starter for client material, health, anything under the data doctrine. Ruled out by physics, not by preference.

Which is where Clef comes in.

Clef is Cloudflare's answer, launched October first, twenty twenty-six, built with TypeSafe, speaking the same System One language. Two differences matter. It runs on Workers A I, Cloudflare's edge network. And its weights are open, under the Apache license. Nobody can take the model away from us.

Two sizes, right?

Flash — nine billion parameters, the sprinter. And the full Clef — twenty-seven billion, the heavyweight. A real decision costs roughly a thousandth of a cent on Flash, at today's prices. We know, because we have been paying it since October fifteenth, inside a lesson we built for someone else.

Why is the edge such a big deal? Fast is fast.

Three reasons. No cold starts: Cloudflare Workers run in lightweight sandboxes that start in roughly zero milliseconds, while a traditional server can take seconds to wake. A decider that answers in forty milliseconds but naps for three seconds is not fast. Second, three hundred plus cities — the model lives near whoever is asking, including here in California. The round trip is a drive across town, not a flight. And third, the memory lives right there too. A database for the ledger, a key-value store for the sticky notes, and Durable Objects when exactly one actor must own a thing.

And this is not theoretical for us.

Not remotely. The Bot Manners rule-check layer has been deciding on this pattern since October seventeenth — three hundred forty-nine logged decisions at last count, and every graded decision in its sample correct. Two honest footnotes. One: that perfect score rests on six graded samples. Promising, not proven. Two: that engine is currently a general chat model wearing the doctrine, at a median of eight hundred forty-two milliseconds a call. The doctrine was proven first. Native Clef is the same contract on a much faster engine. The demo page runs the real native model.

Time for the limitations. All of them.

The three that matter most. First: it can only pick from your menu. Bounded options are the whole design, which means garbage options in, garbage decisions out. If the right answer is not on the list, the decider will confidently pick a wrong one. Designing the menu is the engineering.

Second.

Confidence is not correctness. The number tells you how torn the model is, not whether it is right. A calm zero point nine can still be wrong. Third: it cannot think out loud. No prose, no reasoning, no it depends. Anything genuinely ambiguous, novel, or high-stakes must climb the ladder.

And the work nobody sees?

State design. The model judges the summary you hand it. Leave the decisive fact out of the state, and no model on Earth saves you.

Common things, then. What do people actually do with this?

Route a bark. You dictate a voice note — gym done, remember to text someone, an idea for a widget. One call: which lane first, health, people, ideas, or money. Is there a health report in it. How urgent. Under a second, it is filed, logged, and queued. The demo's second preset does exactly this. Or content approval: one yes-or-no per house rule, tone, a risk score — the clean ninety percent never touches a desk, and the flagged ten percent arrives with the odds attached. Or picking a thumbnail: up to four images can ride along in a call, so the decider can literally judge pictures, in milliseconds, for fractions of a cent.

And the home.

The home operator is my favorite. You hand it sensor readings and the house rules, and ask it to pick one — from the valid actions only. It cannot choose an action that is not on the menu, which is exactly what makes it safe near real switches. Unsure? It does nothing, and logs it.

Now the part everyone actually asked about. Games.

Start with arithmetic. A frame at sixty frames a second lasts sixteen milliseconds. A living game world wants hundreds of little brains ticking. A chat model takes a second just to start answering — you cannot put a novelist inside a goblin. But a decider finishing in about a hundred thirty milliseconds fits the brain tick, the few-times-a-second moment where a character picks its next move.

And games have secretly always worked this way.

Always. The Sims runs on utility A I — every action scored against the Sim's needs, highest score wins. A menu, a scorer, a pick. Clef's exact shape, hand-tuned instead of learned. F E A R made G O A P famous — soldiers with goals, and a menu of actions with preconditions and effects, chained into behavior that looks alive. Same skeleton. Bounded options, chosen fast.

And the world itself?

That is the genuinely new part. Cloudflare ported Doom — yes, that Doom — to run multiplayer in the browser, with Workers routing the game and a Durable Object keeping each session's state. And their engineers published Durable Worlds — a multiplayer voxel game, think small Minecraft, where each world lives inside a single Durable Object. One room, one doorkeeper. Players connect over WebSockets, and the terrain sleeps in storage. Open source, every line readable. Then there was PartyKit, the real-time layer Cloudflare acquired — its hosted service just wound down this month, and the pattern lives on as PartyServer, the same rooms run in your own Worker. A programmer named Tim Bray built a tiny multi-user dungeon on PartyKit, eighty characters, two rooms, turn-based, specifically to prove the game-server pattern. And hibernation: idle connections cost nothing, because a socket can sleep and wake the instant a message arrives. You pay for play, not for waiting.

Now the honest boundary, because I know there is one.

There is. We found no evidence of a shipped commercial game whose characters run on Jev or Clef by name. Anyone who tells you otherwise is selling. What exists is the complete ingredient set, Cloudflare's own playable demos, and enterprise adopters — the Clef launch named Klarna, Salesforce, and Mercado Libre as design partners. The gap between ingredients and a game you would buy is a studio deciding to build it. A when, not an if. But today, still a when.

Before the stopwatch, I want the craft segment. You said designing the menu is the engineering. Teach that.

Gladly. Three rules. Rule one: the options must cover the world. If a bark can be about money, money had better be on the lane list — because the decider will never invent the missing lane, it will squeeze the bark into a wrong one and smile. When we built the routing preset, the lanes came from watching where barks actually go, not from imagining.

Rule two.

Make the options mutually exclusive and boringly clear. Firm and hostile overlap — deliberately, in the email test, because we wanted to see the model wrestle. In production you would sharpen those definitions until the wrestling stops, or you accept the wrestling as signal. Every option can carry a short description with it, and rewriting that description is how you tune the decider. No retraining. Just better words on the menu.

And rule three is the scale.

The scale is an opinion you are allowed to edit. Rude, blunt, neutral, polite, warm — five levels, lowest first. Add a level, rename one, and every future score shifts meaning. That is power and danger in the same knob, which is why our rule system versions its rules: change the wording, stamp the version, and never pretend last month's two point one means the same as this month's.

Give me the failure you have actually seen.

The politeness trap from earlier. Zero point eight four sounds like a good grade until you remember the scale starts at rude. Every team that adopts scoring hits that once — reads the number as a percentage, celebrates, and ships a rude email. The fix is presentation, not math: always show the level the score lands on, in words, next to the number. The demo does exactly that.

And the state? You said state design is the invisible work.

It is the whole game. The decider judges the summary you hand it — nothing else exists for it. When our rule-checker reviews a draft, the state includes the draft and the specific rule being checked, one rule per question, because a state that asks about everything judges nothing well. Short, complete, and honest about what is missing. If a fact is not in the state, the model is not ignoring it. It never saw it.

So the skill stack changes. Less prompt whispering, more menu design.

Exactly. The questions are the program.

Last segment. The stopwatch. These are our measurements, from today, on the live demo.

All three questions in one call — yes-or-no, pick-one, and score together. Flash finished in about a hundred thirty milliseconds. The full twenty-seven billion parameter model, two hundred sixty to six hundred forty. Round trip from this house, tap to answer: about zero point two seconds. Cold first call after quiet: up to two and a half seconds on Flash, five on the big one. And the worst stall we saw, roughly two calls in a hundred, was about twenty seconds — rare, real, and exactly why the fallbacks and circuit breakers exist.

And against a chat model?

Same network, same email question, one of the fast small chat models. It took zero point nine to two full seconds just to produce its first token. And a first token is not an answer — the rest streams in prose, which something then has to read and interpret. Clef's hundred thirty milliseconds is the finished decision, odds attached, ready for a machine to act on. Cloudflare's own published median for Flash is thirty-eight point eight milliseconds — their stopwatch, their best case. Ours says about a hundred thirty in the wild. Believe the wild.

Takeaway.

Most of what a smart operation does all day is not writing. It is judging. Deciders take that ninety-three point eight percent and make it nearly free, nearly instant, and — the part nobody else gives you — honest about their own uncertainty. The craft moves up a level: from writing answers, to designing menus, scales, thresholds, and ladders. Which is, not coincidentally, exactly the business we are in.

The demo is live at clef-lab dot twenty-nine court dot com. Go fire a question at it — and try to make it tear itself in half. That torn feeling? That is the whole lesson.