Skip to content
AI stack, featured

Jev for marketing: Claude + Jev on a real ad market

Last verified

Jev is TypeSafe AI's first System One model: it answers typed questions with probabilities and writes no text. In marketing it is the judge inside a Claude + Jev creative system: Claude describes ads and writes briefs, Jev answers, code counts and a person approves. A real run tagged 54 live ads and pre-flighted six concepts for 1.5 cents.

A typed decision gate
A typed decision gateA sorting machine: example items ride a rail to the position of the probability the model returns, shown on a dial, and drop into one of three bins, block in red, person reviews in amber or pass in green, split by two adjustable blue stops that are the thresholds.00.51
block below 0.30 · pass above 0.70

Drag the slider to move the stops
Jev judges, your code decides, a person approves anything that spends money. Each example item lands at the probability the model returns; the two stops are your thresholds. Widen the band between them and more items go to a person.

Facts

Vendor
TypeSafe AI
Model id as the vendor writes it
jev-1.14.0
Status
Limited access
Access
API at api.typesafe.ai and a web playground. The vendor calls it early access; whether signups are open could not be verified.
Price as the vendor states it
$42 per billion input tokens ($0.042 per million), input tokens only; output tokens free (currently). Source: TypeSafe AI, checked 15 November 2026.
Limits
Default: about 64k tokens for state plus all questions, 32k for state plus one question. Header TypeSafe-Danger-Zone-Long-Context: true raises both to 262k at the same per-token price; vendor says slower and less accurate.
Last verified
15 November 2026

The system in three steps

Three agent skills split one creative cycle into parse, brief and pre-flight. Claude does the reading and the writing. Jev answers narrow typed questions with probabilities. Code counts, applies thresholds and re-checks numbers, and a person approves anything that leads to production or spend.

  1. 1.Parse the market

    Claude looks at each ad from a public ad library and writes a plain description. Jev tags each description with Choice and yes or no questions from a taxonomy drafted for that market. Code counts the tags by track and by advertiser into a market snapshot and sends low-confidence tags to a review queue. Skill: competitor-creative-parser.

  2. 2.Write the briefs

    Claude reads the snapshot and the team's own rules, picks the moves and writes two brief packs: one for static designers and one for the motion team, with three hooks per video concept. Code re-reads each cited number from the snapshot before the pack renders. Skill: market-to-creative-brief.

  3. 3.Pre-flight the concepts

    Claude writes 12 or more example buyer profiles from the team's notes and the snapshot. Jev answers five typed questions for each concept and profile pair. Code builds a heat map and returns go, rework or drop with reasons, and a person picks what goes to a live test. Skill: creative-preflight-panel.

Who does what in the Claude + Jev creative system
RoleDoesDoes not
ClaudeDescribes ads, drafts the market taxonomy, writes decisions, briefs and example profilesCount shares or decide what launches
JevAnswers Choice, yes or no and Score questions with probabilitiesSee images or video in this run, or write text
CodeCounts, applies thresholds, re-checks cited numbers, builds the review queue and the heat mapJudge meaning
A personClears the review queue, approves the brief pack, picks concepts for the live testHand a budget to the panel

The real run: 54 live ads, 1.5 cents

The run read 54 active ads for AI photo editor apps from 23 advertisers in the public Meta Ad Library, US, captured on 15 November 2026: 20 static and 34 video. Jev ran as jev-1.14.0 on the vendor's direct endpoint. The advertisers stay anonymous on this page, with no page names, no ad ids and no copy from their ads. The team that used the result, Brand Z, is an example, and so are its six concepts and the 12 buyer profiles.

Ledger of the final run, Jev input tokens at the listed price
StepRequestsTyped answersInput tokensUS dollars
Tagging, 54 ads1241,371204,5040.0086
Panel, 144 concept and profile pairs168720142,0220.006
Final run, total2922,091346,5260.0146
Four pilot rounds before it304,3590.0128

The dollars cover Jev input tokens at $0.042 per million, and output tokens were free at the time of checking: 346,526 tokens times $0.042 per million is $0.0146. Claude's own work ran in a chat session and sits outside these numbers. Code accepted 222 of 304 Choice tags without a person, 73%, and sent 107 rows to the review queue. Asked again with the options in reversed order, 85 of 88 Choice answers came back the same.

Static and video ads in this market follow different playbooks. The static ads sell with an offer and proof, and the videos show the transformation and ask for little.

  • Static, 20 ads: professional headshots fill 13 of 20 creatives and 4 of 7 advertisers. 12 of 20 name an offer, 8 a free preview and 4 a free start. 11 say free, 10 cite social proof, 4 show a price and none uses urgency.
  • Video, 34 ads: 29 of 34 state no offer, 1 shows a price, and none cites social proof or uses urgency. At least 12 are fully AI generated clips, and 25 put text on screen in the first three seconds.
  • Proof of the result in video: 10 use a small inset of the source photo, and 1 shows the original and the result at equal size.
Ad market x-ray replay
Live ads read
5420 static, 34 video, 23 advertisers
Typed answers from Jev
2,091292 requests, 346,526 input tokens
Jev cost, final run
1.5 centsUS$0.0146 for the final run

Static against videocreatives that do it, of each track

Static against video: creatives that do it, of each track
PatternStatic, 20 adsVideo, 34 ads
Names an offer12 of 205 of 34
Cites social proof10 of 200 of 34
Shows a price4 of 201 of 34

Pre-flight panelchance an example buyer stops or clicks, before production

Pre-flight panel: each row is an example concept or a hook of one, each column an example buyer profile, each cell the chance that profile stops or clicks. Profile 12 is the control. The last two columns are the row mean and the verdict code gave.
ItemProfile 1, Job seeker, profile photo this weekProfile 2, Freelance consultant, sends proposalsProfile 3, Real estate agentProfile 4, Team lead, remote team pageProfile 5, Designer, has seen every AI headshot adProfile 6, Student, tight budgetProfile 7, Privacy-minded professionalProfile 8, Shares fun AI clips with friendsProfile 9, Refreshing a dating profileProfile 10, Paid before, results looked wrongProfile 11, Coach building a personal brandProfile 12, no need for portraits, controlMeanVerdict
Static, 3 concepts, 3 items
S2 Photo shoot from selfies.96.88.71.81.58.90.64.88.98.99.99.110.79rework
S3 Price as the headline.90.94.87.67.70.91.75.37.65.99.88.100.73go
S1 Headshot for freelancers.96.97.70.60.80.82.85.20.42.99.66.010.67rework
Video, 3 concepts, 9 items
M1-C See it before you pay hook C.90.88.77.76.88.95.72.84.89.98.89.220.81go
M2-B Before and after, same size hook B.94.77.73.94.92.84.57.92.88.95.92.080.79go
M1-A See it before you pay hook A.80.85.80.76.73.97.56.77.78.99.97.070.75rework
M1-B See it before you pay hook B.78.89.71.83.46.87.60.47.84.99.95.080.71rework
M2-C Before and after, same size hook C.92.68.59.91.95.55.25.71.78.93.88.040.68rework
M2-A Before and after, same size hook A.86.73.60.94.92.78.30.60.57.95.84.050.68rework
M3-A Real person reacts hook A.85.93.75.72.40.84.75.18.74.78.65.010.63rework
M3-B Real person reacts hook B.81.72.52.54.83.79.52.52.63.88.57.100.62rework
M3-C Real person reacts hook C.75.47.46.33.68.74.47.48.72.96.54.080.56rework

Each cell is the chance that an example buyer stops or clicks. Profile 12 is the control, a person with no need for portraits.

Example buyer profiles

  1. 1 Job seeker, profile photo this week
  2. 2 Freelance consultant, sends proposals
  3. 3 Real estate agent
  4. 4 Team lead, remote team page
  5. 5 Designer, has seen every AI headshot ad
  6. 6 Student, tight budget
  7. 7 Privacy-minded professional
  8. 8 Shares fun AI clips with friends
  9. 9 Refreshing a dating profile
  10. 10 Paid before, results looked wrong
  11. 11 Coach building a personal brand
  12. 12 no need for portraits (control)
Hover, focus or tap a cell
A replay of one real run. The ads are live ads from a public ad library, the advertisers stay anonymous, and the team, its concepts and the buyer profiles are examples. A panel score filters concepts before production; the live test decides.

One input and one output from the run

Each heat map cell comes from one request. For the cell of concept S3 and example profile 4, the state held the concept as written for the designers, a text-led card with the price as the headline and a free preview line, plus the profile of a team lead who runs a remote team page. The request used 787 input tokens. Jev returned five typed answers, and code read them as below.

A real panel request: concept S3 shown to example profile 4
QuestionTypeJev's answerCode reads it as
action: the buyer's most likely move on seeing the adChoice: scroll, stop, clickstop 0.23, click 0.44, scroll 0.33, confidence 0.16Stop or click 0.67, the heat map value; inside the review band
hook_about_problem: the headline is about this buyer's problem or goalYes or no0.47Review band, not counted as yes
offer_clear: the offer and its terms are clear at first sightYes or no0.52Review band, not counted as yes
distrust_claim: this buyer would doubt a claim in the adYes or no0.43Review band, not counted as distrust
fit: how well the concept fits this buyerScore, 0 to 32.31, confidence 0.52Feeds the concept's mean fit; a go needs 2.0

Across the panel, code gave go to S3, to M1 with hook C and to M2 with hook B, and rework to S1, S2 and M3. S2 has the highest static mean and still gets rework: a third of the profiles would distrust a claim, and none found the offer clear at first sight. The control profile, a person with no need for portraits, scored 0.01 to 0.22 on each item, so the panel separates buyers from non-buyers on this run. That checks the panel, and says nothing about how good the ads are.

What broke first, and the two fixes

The first question set was written for a made-up market, and it failed on real ads. On the first 31 ads it sent 143 rows to review, and in video 12 to 22 of 22 creatives stayed unresolved per dimension. An ad library capture has no audio, and the generic options did not fit a market built around what an app does with a selfie.

  • A market taxonomy. Claude drafted the options from the descriptions of the first 30 ads: five to ten options per dimension, other last, and questions that read only fields the capture has.
  • A confirm pass. Each tag gets a second, narrower yes or no question about the chosen option, which replaces the generic companion question as the second vote.

On all 54 ads the review queue fell to 107 rows, and 222 of 304 Choice tags resolved without a person. The fixes landed together, so this run cannot say which one mattered more. The pilots cost almost as much as the final run, $0.0128 against $0.0146, so budget two or three pilot rounds before the full pass.

Run it yourself

  1. 1.Install the three skills

    They ship in the user-acquisition-skills pack, 15 skills under the MIT licence. Run them in order: parser, brief, panel.

    npx skills add ua5er/user-acquisition-skills --skill competitor-creative-parser
    npx skills add ua5er/user-acquisition-skills --skill market-to-creative-brief
    npx skills add ua5er/user-acquisition-skills --skill creative-preflight-panel
  2. 2.Put the key in the environment

    The scripts read TYPESAFE_API_KEY from the environment and from nowhere else, so the key stays out of chat, files and logs. Keys come from the TypeSafe console. Access is early access, and the homepage does not say whether signups are open.

    export TYPESAFE_API_KEY="<your key>"
  3. 3.Start offline, then cap the spend

    A dry run uses an offline mock, a smoke run sends three live pairs to check the response shape, and a spend cap stops the full run. Send text descriptions to Jev; in this run images and video stayed with Claude.

    python3 scripts/tag_creatives.py descriptions.jsonl --track static --dry-run --out-dir run
    python3 scripts/tag_creatives.py descriptions.jsonl --track static --smoke --data-handling-confirmed --out-dir run
  4. 4.Know the price and the limits

    The docs list $0.042 per million input tokens, with output tokens free at the time of checking. Default context is about 64k tokens for the state plus all questions and 32k for the state plus one question; a long-context header raises both to 262k. The body limit is 32 MiB, and rate limits are set per organization and not published in the docs.

Try the pre-flight on one ad of your own. Paste the text and get typed checks, policy flags and counted facts in seconds, with no signup.

Open the free ad creative checker

What this run does not prove

  • An ad library shows what runs, never spend or results. A gap means nobody in this sample stands there, and it says nothing about whether the idea sells.
  • The tags are model output with a review queue. The 73% that code accepted without a person is a share and carries no accuracy: nobody labelled a sample of these ads.
  • The thresholds, tag at 0.70 and review from 0.35, are the vendor's starting values and were not calibrated on labelled ads.
  • The sample is 54 creatives from one day, newest first, with no audio.
  • A panel score is a filter and a signal. The profiles are examples written by Claude, the panel touches no budget, and the live test decides.

Questions

What is Jev and what does it return?
Jev is TypeSafe AI's first System One model. You send a state and named questions, and each question comes back typed: a Noul returns a probability from 0 to 1, a Choice returns probabilities for every option plus a confidence, and a Score returns a weighted level plus a confidence. Box is experimental. Jev does not generate text.
What does Jev do in the Claude + Jev creative system?
Jev answers the typed questions. It tags each competitor ad description with Choice and yes or no answers, and in the pre-flight panel it answers five questions for each concept and profile pair. Claude writes the descriptions, briefs and profiles, code counts and applies the thresholds, and a person approves.
How much did the real run cost?
The final run cost $0.0146 in Jev input tokens: 292 requests and 346,526 input tokens at $0.042 per million, with output tokens free. Four pilot rounds before it cost $0.0128. Claude's own work ran in a chat session and is not in these numbers.
Does Jev see the ads themselves?
Not in this run. Claude looked at the images and video frames and wrote text descriptions, and only those descriptions went to Jev. The vendor lists image input for jev-1.14.0, which this run did not use.
Can a panel of example profiles replace a live test?
No. The profiles are examples written by Claude, and a panel score is a cheap filter before production. It cut three of six concepts to rework here, and the survivors still go to a live test that decides.
How do I get access to Jev, and is the waitlist still on?
API keys and a no-code playground are in the TypeSafe console. The homepage says early access and does not say whether signups are open, and third-party reports conflict, so check at signup time. Jev is also listed on Vercel AI Gateway as typesafe-ai/jev.
Is it safe to send customer or ad data to Jev?
The docs say Jev is not trained on customer requests or responses, zero data retention is available to enterprise customers on request, and a Data Processing Agreement exists. Retention periods are not stated and the pages do not address ad or customer data specifically. On Vercel AI Gateway you can select Zero Data Retention per request. Strip personal data first.

Re-check log

  • 15 November 2026: rebuilt around a real run with jev-1.14.0 on the direct endpoint: 54 live ads tagged and six example concepts pre-flighted, with the ledger above. Price, limits and access come from the vendor docs as checked the same day.
  • Next re-check: within 30 days, because the price, the model id and the access terms can change.

Sources

Each source shows the date it was last checked.

Show all 17 sources
  1. TypeSafe AI docs.typesafe.ai/agent-skill.md
  2. TypeSafe AI docs.typesafe.ai/api/system-one.md
  3. TypeSafe AI docs.typesafe.ai/changelog.md
  4. TypeSafe AI docs.typesafe.ai/confidence.md
  5. TypeSafe AI docs.typesafe.ai/cookbooks/llm_guardrails.md
  6. TypeSafe AI docs.typesafe.ai/introduction/quickstart.md
  7. TypeSafe AI docs.typesafe.ai/legal.md
  8. TypeSafe AI docs.typesafe.ai/model-jaggedness.md
  9. TypeSafe AI docs.typesafe.ai/models.md
  10. TypeSafe AI docs.typesafe.ai/patterns.md
  11. TypeSafe AI docs.typesafe.ai/primitives.md
  12. GitHub github.com/typesafe-ai/skills
  13. GitHub ua5er/user-acquisition-skills with the Claude + Jev skills
  14. TypeSafe AI typesafe.ai
  15. TypeSafe AI typesafe.ai/blog/dont-call-it-a-decision-model
  16. TypeSafe AI typesafe.ai/blog/introducing-system-one-models-and-jev
  17. GitHub Example: the Claude + Jev system on a real market

Want an AI judge wired into your ad workflow with approval gates? Start with a short audit.

See the service