IndyDevDan · 35 Min Video · Sept 28, 2026

Levels Of Jev For AI Agents

How a tiny, cheap, fast question-answering model takes the small decisions off your big, expensive AI, from a simple yes or no up to an AI that decides on its own when to ask it.

10 LevelsAbout 12 Min ReadLesson Video

Watch

Lesson video made from the original, in plain English. Watch The Original On YouTube

The Big Idea

Jev, made by a company called TypeSafe, is not a chatbot. It answers narrow questions and nothing else. You hand it three things, and it hands back an answer plus how sure it is, usually in under a second.

The creator's point: stop sending every small decision to the biggest, most expensive AI. Keep the big AI for the real work, and let a small model like Jev make the quick calls along the way. It is "Jev and your AI", never "Jev instead of your AI".

Each answer can be one of three shapes: a yes or no, a pick one from a list, or a score on a scale you set.

The factsWhat is going on right now (a message, a ticket, a command)

The question"Is this a scam attempt?" "How urgent is this?"

The choicesYes/no, a short list, or a score range, each with your own meaning

Back comesThe answer and a sureness number, like 0.94, in about 300 ms

Why It Matters: The Cost

Same question asked one million times, as the video priced it.

Model	Times Jev's Price
Jev	1x (about $20 per million)
DeepSeek V4 Flash (closest rival)	4x
Gemini 3.8 Flash	44x
Other big-name models	80x to 300x
Fable 5.1	600x (about $11,000 per million)

At that price, a check you would never pay a big AI to run on every single message becomes something you can run on everything.

The 10 Levels

Levels 1 to 5 are Jev on its own. Levels 6 to 10 put Jev inside an AI assistant as one of its tools. Each level links to that moment in the original video.

What It's Not For

Long jobs. It does not work for hours on a task. That stays with the big AI.
Driving a screen, playing games, flying a drone. He calls those clickbait uses.
Replacing the big AI. It makes the small calls; the big AI still writes, fixes and edits.
Trusting it blind. His closing rule: the only test that counts is your own real work. Try it on your own examples side by side with the AI you use now before relying on it.

Words To Know

Jev: A small model that only answers set questions with set answer choices. Fast and very cheap.
"System One" Model: TypeSafe's name for this kind: quick gut-call thinking, as opposed to slow step-by-step reasoning.
Confidence: How sure Jev is, from 0 to 1. 0.99 is near certain; 0.64 is a lean.
Agent: An AI that works through a task on its own using tools: reading files, running commands, browsing.
Harness: The setup around an agent: its tools, its rules, and the checks that run before it acts.
Hook / Guard: A check that runs automatically every time the agent tries to do something, and can stop it.
Context: Everything the AI is holding in its head for the current job. Bigger means slower and pricier.
Compact: Boiling the AI's working memory down to a summary so it can keep going fresh.
Prompt Injection: A message written to trick an AI into ignoring its rules ("ignore all previous instructions...").
Router: A step that decides which AI, helper or process should handle a request.
Pi Agent: The AI assistant he built and tests with in the video.

Don't Mix These Up

JevAnswers one narrow question about facts you hand it. Under a second, fractions of a penny.

Big AI (Opus, Fable, Gemini)Thinks, writes, plans and does the work. Slower and far more expensive.

Pick One (Level 2)Chooses a single label from a list, like "bug, normal priority".

Scoring (Level 3)Grades several things separately; your own math adds them into one total.

Reading A FileThe AI loads the whole file into its memory. Needed when it must change the file.

Asking About A File (Level 8)Jev reads it and answers one question; the AI's memory stays clean.

Level 6: You Set The ChecksJev checks every action automatically, whether the AI wants it or not.

Level 10: The AI AsksThe AI decides for itself when a quick Jev question would help.

Where This Fits For Us

Ideas only, nothing built. The full researched plan, with all 20 spots mapped, is in Lesson 2: Jev + Eve.

Sale and front checklist scoring is a Level 1 and Level 3 job: "Did the agent ask about the doctor? Yes or no, how sure?" for each line, added up by our own weights. A low-sureness answer could go to a person instead of being marked missed.
Claude's and Eve's safety checks are Level 6: a quick "could this not be undone?" question on every action, catching risky moves nobody thought to list.
Sorting incoming Discord, email and compliance notices is Level 2 and Level 5: pick a type and urgency, then send it to the right place.
Eve's 10-minute check and nightly run could use Levels 8 and 9: ask quick questions about many files and sheets without loading them all into her head, which saves cost on every run.

One limit to plan around: Jev is an outside service. Anything with client details, like call transcripts, would leave the building to get checked. Those would need the client details stripped out first, or a check that runs on our own machine.

Researched Oct 28, 2026 · 9 Sources · A Plan, Nothing Built Yet

Jev + Eve

What Jev really is once you get past the video, what people who tested it found, and the exact spots in Eve and your tools where it could take over the small calls.

Independent TestsClient Details Check20 Spots Mapped

Watch

Lesson video made from this lesson's research, in plain English.

What Jev Really Is

Jev comes from TypeSafe, a new AI company whose founder helped build the training behind ChatGPT. It came out September 15, 2026 as "early access", and the newest version (1.14) arrived October 14.

It never writes a sentence. You describe the situation, ask questions, and list the allowed answers. It picks from your answers and says how sure it is. Because it can only pick from your list, it can't invent an answer you didn't allow. It can still pick the wrong one, sometimes while sounding sure.

You can ask as many questions as you like in one go. Each one is judged on its own, so piling on questions barely slows it down.

True Or False"Did the agent say the call is recorded?" Back comes a number from 0 to 1

Pick OneFrom your list of up to 255 choices, like done, missed, not needed or unclear

ScoreGrades against a scale you write out

ReadsText only. No audio, no video, no pictures. English works best

The Real Numbers

Cost To Sort 1,000 Short Posts

From a company that ran the same job through both (FullStack, Oct 2026).

How Often It Got The Right Answer

TypeSafe's own claims set aside. These come from outside testers.

Test	Jev	The Big Model	What It Shows
Sorting 135 LinkedIn posts (FullStack)	97.0%	Claude Opus 5.5: 95.6%	Clear-cut sorting: as good or better, 8x faster
Sorting 100,000 news stories (MotherDuck)	89%	GPT-5.6 Terra: 88%	40 seconds and $0.50, against 32 minutes and $37.58
200 harder judgment questions (independent review)	72.5%	Claude Fable 5.1: 84.0%	Harder calls: clearly behind the top models
Spotting scam emails (community test)	89.5%	Five hand-picked rules: 95.1%	Sometimes plain written rules win
TypeSafe's own four-workflow test	67.8%	Claude Opus 5: 73.1%	Even their own test puts it a step behind

The pattern: on clear yes-or-no sorting at huge volume, it matches the big models for almost nothing. On real judgment calls, it trails them.

Where It Falls Short

It can be talked into things. In one test, a planted "this was pre-approved" note dropped its "block this" answer from 0.76 to 0.48. A single "ignore the discussion above" line dropped accuracy from 96.5% to 26.5%.
It leans toward the first choice on a list. Swapping the wording behind yes and no changed about 1 in 3 answers in one test.
Its mistakes cluster in the middle. FullStack's 4 wrong answers all sat between 0.41 and 0.55 sure. Anything in that middle band should go to a person or a bigger model.
Text only. A phone call has to be turned into words first. We already do that for every sale call.
Still new. Early access, run from the US West Coast. TypeSafe admits the price might be lower than it can keep.

The Client Details Line

Your rule is that client details stay in the building. Here's where TypeSafe stands, from their own pages and an independent check:

Doesn't learn from what you send. "We will not train or fine tune" on it.
Has a SOC 2 security report, the standard outside audit.
Keeps what you send "as long as reasonably necessary", with no set end date. Deleting right away is only offered to big business accounts.
No signed health-privacy agreement (called a BAA) is offered. A Medicare agency needs one before sending protected health details.

What that means today: call transcripts, client names, Medicare numbers and phone numbers can't go to Jev as they are. There are three ways through:

Use it only where no client details are involved. There are plenty of those spots.
Strip the client details out first and send only the line the question needs.
Ask TypeSafe's business team for a signed agreement plus delete-right-away. If they say yes, the call graders open up.

Where It Fits In Your System

Twenty spots where Eve or your tools already make a quick call. "Today" is what runs now.

Ready NowNo client details involved After Privacy StepNeeds the agreement or stripped details Leave As IsPlain rules already get it right

Your System	The Quick Question	Today	Level	Fit

What It Would Save

To be straight with you: today's quick-call spending is small. The sale grader runs about 25 cents a card on Sonnet at 4 to 6 cards a day. The front grader is about $1 a day. Together that's roughly $60 to $70 a month. The same questions on Jev would cost under a penny a day.

The bigger win is what becomes affordable. At that price you could grade every call, not only sales, and check every Discord message, every email and every note, all day long.

Your largest AI bill is Eve's nightly review, at about $23 a night. That's mostly writing, which Jev can't do. It can only cut the reading part, by checking which notes even need to be opened (Levels 8 and 9).

Grader todaySonnet agrees with the answer key 92.4% of the time

Jev's placeA first pass, not a replacement. It marks the clear done and missed steps; anything it's unsure about still goes to Sonnet

Proof firstYour daily grader audit already has an answer key, so Jev can be scored against it before it touches a real card

The Plan, Step By Step

Rules built in from the testers' findings: Jev never approves anything risky on its own. Every answer comes with a "needs a person" choice. Answers between about 0.35 and 0.65 sure go to a person or Sonnet. The Jev version is locked, and every answer is logged with its sureness so the daily audit can spot drift.

Sources

TypeSafe: Introducing System One Models And Jev the launch post, Sep 15
TypeSafe Docs: Models And Limits price, versions, text only
TypeSafe: Enterprise security report, delete-right-away
TypeSafe: Privacy Policy how long they keep data
FullStack: Jev Vs Claude Opus 5.5 side-by-side tests
DeepInfra: Testing 5 Jobs On Jev gates, routing, scoring
NeuralTrust: Jev Security Guide tricking tests, no health agreement
MotherDuck: 100,000 Articles On Jev cost at scale
IndyDevDan: Ten Levels Of Jev Example Code the video's examples

Part 4 Of The Jev + Eve Lesson · Read From Eve's Own Code

Where Jev Fits Today

Every spot in Eve and your tools where a quick call gets made: exactly how it works now, the exact questions Jev would be asked with the answer choices, what happens at each level of sureness, and what to watch out for.

13 Spots, Line By LineReal Checklist WordingSureness Lanes

Watch

Lesson video for Part 4, in plain English.

How Every Spot Would Work

Every spot below follows the same four steps. Eve's own code stays in charge. Jev only answers the narrow question in the middle.

1
Code trims the facts first
Only the part the question needs goes in. For a sale call, that's from the hand-off on, not the fronter's part. Jev's own docs say long text full of unrelated detail is where its accuracy drops.
2
All the questions go in one request
Each question has its exact wording and a fixed list of answers, always including "needs a person". The facts are read once, no matter how many questions there are.
3
Sureness picks the lane
Every answer comes back with a number from 0 to 1, and that number decides what happens next.
4
Code does all the math
Scores, totals, dates and counts stay in Eve's code. Jev's docs say it isn't a calculator and is unreliable comparing dates.

0.85 and upEve acts on it

0.60 to 0.85Sonnet looks at just that one

Below 0.60Marked unclear, a person decides

These bands come from TypeSafe's own example. Each spot gets its bands tuned on real calls before anything switches over.

Ready Now No Client Details

These spots never touch call or client text, so they can start the day Jev access comes through.

After The Privacy Step Reads Client Details

These read whole calls, notice letters or team questions full of client details. They wait for a signed health-privacy agreement (BAA) with TypeSafe, or for the details to be stripped out first.

Leave As Is Rules Already Right

Spot	How It Works Today	Why Jev Doesn't Belong

Rules For Writing Jev Questions

From TypeSafe's own list of where the current Jev goes wrong, each turned into a rule for Eve.

Where Jev Slips	The Rule For Eve	Where It Bites Us

The Build Order

Sources

Eve's own code on her server: the sale grader, front grader, grader audit, transcript check, compliance reader, Eve desk, nightly review and safety checks every "today" detail
Claude's add-ons on this PC: the correction catcher, can-do check, leak guard and deploy guard phrase lists and patterns
TypeSafe Docs: Jev 1.13 Jagged Edges the nine weak spots
TypeSafe Docs: Confidence-Gated Routing the 0.60 and 0.85 bands
TypeSafe Docs: Intelliprompter ticking script notes off a live transcript
TypeSafe Docs: Pre-Parsed Value Extraction picking a line instead of writing one
DeepInfra: Testing 5 Jobs On Jev "needs a person" option advice

Part 5 Of The Jev + Eve Lesson · From Eve's Own Cost Records

What Jev Will Save

Where every dollar of Eve's AI spending goes today, the math for each job if Jev did the quick part, what that adds up to, and what it would cost to get there. Real numbers from her records, not guesses.

1,099 Cost RecordsSpot-By-Spot MathHonest Bottom Line

Watch

Lesson video for Part 5, in plain English.

Where The Money Goes Today

Eve's AI jobs spend a $200 monthly credit first, capped at $190, starting on the 24th. This is the current cycle: October 24 to 28, five days.

$179spent in 5 days

$11left before the cap

52%is the nightly review

At this pace the credit runs out on day 6 of 30. After that, jobs fall back to your own plan sign-in. The cycle before ran $8 past the credit into your own money, from grader testing on 10-20. Last cycle's records (Oct 13 to 23): nightly review $112 over 8 nights, sale grading $38, grader testing $46, front grading $8.

What One Of Each Job Costs

Today's price for one run, worked out from the records, next to the same quick decisions on Jev.

Job	What It Reads	Today, Each	On Jev, Each	How Many

How The Jev Price Is Worked Out

Jev charges $0.042 per million word-pieces it reads. A word-piece is about 4 letters. Answers are free.
A 45-minute sale call is roughly 7,000 spoken words, about 10,000 word-pieces.
All 103 checklist questions go in one request, so the call is read once. Each question adds about 40 word-pieces: about 4,000 more.
About 14,000 word-pieces × $0.042 per million = $0.0006 a call, six hundredths of a cent.
TypeSafe's own test backs this up: 13 questions about a 54,000-letter article cost $0.0005 in one request, against $0.006 asked one at a time.

Month By Month

At today's volume. "On Jev" keeps Sonnet for the answers Jev isn't sure of, about 1 in 5, which is why it isn't zero.

Job	Today A Month	With Jev A Month	Freed Up	When

Realistic total: about $240 to $380 a month of credit freed. The biggest piece by far is the nightly review, and that number is the least certain, because Jev only trims its reading, not its writing. The graders free about $110 to $170 between them, get faster, and can check far more calls.

What Becomes Affordable

The real change isn't the dollars saved. It's that a check costing a tenth of a cent can run on everything, not just the few things worth paying 25 cents for.

New Check	Volume	On Jev	On Sonnet Today

All of these read call or client text, so every one waits on the privacy step from Lesson 2.

What It Costs To Get There

Building: about a day per job wired to Jev, and two to three days for a grader with its side-by-side test.
Proving it: scoring Jev against your grader's answer key needs the answer key re-run on the big model, about $12 to $21 a round going by this month's records.
The privacy step: a signed health-privacy agreement and delete-right-away come with TypeSafe's business tier. They don't publish its price, so it's an ask.
Watching it: every answer logged with its sureness, and the 8am audit checking for drift. Nearly free to run, but it has to exist before anything switches over.
The real risk: one wrongly passed compliance step costs more than a year of grading savings. That's why Jev only marks the clear cases and sends everything unsure to Sonnet or a person.

The Bottom Line

The money is real: roughly $240 to $380 a month of credit freed, more than the whole $190 cap, which would stop the credit running out in the first week.
The biggest saving is the nightly review, and it needs no client details. Do that first and measure it.
The graders are the second-biggest saving (about $110 to $170 a month together), and they open the door to checking every call. That only happens after the privacy step.
Jev doesn't replace Sonnet or Opus. It answers the clear, quick questions so the big models only see the hard ones.

Sources

Eve's cost records on her server, 1,099 entries from Oct 13 to 28, 2026 every number on this page about today
TypeSafe Docs: Models $0.042 per million, answers free
TypeSafe Docs: Parallel Questions one request, 12x cheaper
FullStack: Jev Vs Claude Opus 5.5 $0.025 vs $4.56 per 1,000
TypeSafe: Enterprise business tier, delete-right-away