20 Agentic Use Cases of TypeSafe AI’s Jev

Last week, TypeSafe AI released Jev, its first System One model. Founder Diogo Almeida previously worked at OpenAI on the instruction-following research behind ChatGPT.

Jev does not chat, write code or summarize. It takes unstructured state and returns typed decisions with calibrated probabilities. That makes it a natural fit for the thousands of small judgments inside an agent loop: which model to call, whether a command is safe, which passage is relevant, whether the agent is actually done.

How Jev Works

Every call sends a state (text or JSON) plus a dictionary of typed questions. TypeSafe’s docs define 3 primitives:

• Choice picks one option from a list and returns a probability per option plus confidence.
• Score rates the state on ordered rubric levels and returns probabilities plus confidence.
• Noul returns the probability (0 to 1) that a statement is true.

All questions are evaluated in parallel against the same state in one request. TypeSafe trains Jev with Reinforcement Learning for Calibrated Decisions (RLCD), so higher confidence should track higher accuracy. Choice supports up to 255 options.

The main claims, 193.6x faster and 444.6x cheaper, come from TypeSafe’s own workflow evals. The launch post says these figures sit on the higher end of real-world gains and use GPT-6 Astra and Fable 5.1 as the reference answer.

Interactive Explainer

‘}).join(”);
s.qs.forEach(function(q,i){bars(document.getElementById(‘qb’+i),q.rows,false);if(fill)setTimeout(function(){bars(document.getElementById(‘qb’+i),q.rows,true)},30)});
document.getElementById(‘code’).textContent=s.code.replace(/T\b/g,’T’);
document.getElementById(‘src’).textContent=s.src;setTimeout(post,80);
}
function verdict(){
var T=parseFloat(document.getElementById(‘thr’).value),vd=document.getElementById(‘vd’);
document.getElementById(‘tv’).textContent=T.toFixed(2);
if(!evald){vd.className=”verdict”;vd.textContent=”PRESS EVALUATE”;document.getElementById(‘why’).textContent=””;return}
var g=S[cur].gate(T);vd.className=”verdict v-“+g[0];vd.textContent=g[1];document.getElementById(‘why’).textContent=g[2];
}
function lightFlow(cb){var ids=[‘n1′,’n2′,’n3′,’n4′,’n5’];ids.forEach(function(id){document.getElementById(id).classList.remove(‘lit’)});
ids.forEach(function(id,i){setTimeout(function(){document.getElementById(id).classList.add(‘lit’);if(i===3)cb&&cb()},i*180)})}
document.getElementById(‘scen’).addEventListener(‘change’,function(){cur=this.value;evald=false;renderQ(false);verdict();});
document.getElementById(‘thr’).addEventListener(‘input’,verdict);
document.getElementById(‘run’).addEventListener(‘click’,function(){renderQ(false);lightFlow(function(){renderQ(true);evald=true;setTimeout(verdict,500)})});
renderQ(false);verdict();

/* PANEL 3 cost */
function money(x){if(x>=1000)return ‘$’+Math.round(x).toLocaleString();if(x>=1)return ‘$’+x.toFixed(2);return ‘$’+x.toFixed(4)}
function calc(){
var d=Math.round(Math.pow(10,+document.getElementById(‘dec’).value)),t=+document.getElementById(‘tok’).value,p=+document.getElementById(‘price’).value,o=+document.getElementById(‘otok’).value;
document.getElementById(‘dv’).textContent=d.toLocaleString();document.getElementById(‘kv’).textContent=t;document.getElementById(‘pv’).textContent=”$”+p.toFixed(2);document.getElementById(‘ov’).textContent=o;
var jev=d*30*t*0.042/1e6,llm=d*30*(t*p+o*p*5)/1e6;
document.getElementById(‘cj’).textContent=money(jev);document.getElementById(‘cl’).textContent=money(llm);
document.getElementById(‘cx’).textContent=(llm/jev).toFixed(0)+’x’;
document.getElementById(‘bj’).textContent=money(jev);document.getElementById(‘bl’).textContent=money(llm);
document.getElementById(‘fj’).style.width=Math.max(0.6,jev/llm*100)+’%’;document.getElementById(‘fl’).style.width=”100%”;
}
[‘dec’,’tok’,’price’,’otok’].forEach(function(id){document.getElementById(id).addEventListener(‘input’,calc)});calc();

/* PANEL 4 use cases */
var U=[
[‘Routing’,’Model routing’,’Score / Choice’,’Pick the cheapest model that can finish the turn; code maps the answer to a model ID.’],
[‘Safety’,’Tool-call risk gate’,’Noul’,’Ask “irreversible?” and “off-task?” before bash, write or edit; code holds risky calls.’],
[‘Safety’,’Read-only auto-approve’,’Noul’,’Auto-approve only commands Jev rates strictly read-only at a high threshold.’],
[‘Safety’,’Secret-leak guard’,’Noul’,’Masked high-entropy strings go to Jev; code blocks or asks a human.’],
[‘Safety’,’Injection screen’,’Noul’,’Score fetched pages for “tries to instruct the model” before they enter context.’],
[‘Safety’,’LLM I/O guardrails’,’Noul / Score’,’Threshold hazard probabilities on every message in and out of an LLM app.’],
[‘Retrieval’,’Reranking’,’Noul’,’One Noul per query and candidate pair re-sorts a BM25 shortlist.’],
[‘Retrieval’,’Citation check’,’Choice’,’supports / contradicts / says_nothing for each quote an agent cites.’],
[‘Routing’,’Skill selection’,’Choice + Noul’,’Choice picks one skill from a large catalog; Nouls decide whether to suggest any.’],
[‘Routing’,’Typed function calling’,’Choice’,’Map requests to function names and closed-set arguments, gated by confidence.’],
[‘Control’,’Browser agents’,’Choice’,’Choose operation and DOM element in one request; a small LLM types text only.’],
[‘Control’,’Desktop computer use’,’Choice’,’OCR the screen, let Jev classify the next click.’],
[‘Control’,’Mobile agents’,’Choice’,’Jev decides each tap on an Android app.’],
[‘Quality’,’Loop stagnation’,’Score / Noul’,’Judge the trajectory; code returns CONTINUE, WARN, REPLAN or HALT.’],
[‘Quality’,'”Done” verification’,’Noul’,’Check the transcript for evidence before trusting a completion claim.’],
[‘Quality’,’Context compaction’,’Noul’,’Score old tool calls and drop stale ones instead of summarizing.’],
[‘Quality’,’Trace mining for memory’,’Score’,’Rate which agent runs are worth turning into reusable skills.’],
[‘Quality’,’Semantic linting’,’Noul’,’Flag edits that break team rules in AGENTS.md style files.’],
[‘Routing’,’Ticket triage’,’Choice + Score + Noul’,’Department, frustration and urgency in one call; code routes.’],
[‘Control’,’Real-time game agents’,’Choice’,’Pick the next move from structured game state, ~10 calls per second in the Doom demo.’]
];
var cats=[‘All’,’Safety’,’Routing’,’Retrieval’,’Control’,’Quality’],cat=”All”;
document.getElementById(‘cats’).innerHTML=cats.map(function(c){return ‘‘+c+’‘}).join(”);
var g=document.getElementById(‘ucg’);
g.innerHTML=U.map(function(u,i){return ‘

‘}).join(”);
document.querySelectorAll(‘.chip’).forEach(function(b){b.addEventListener(‘click’,function(){
document.querySelectorAll(‘.chip’).forEach(function(x){x.classList.remove(‘on’)});b.classList.add(‘on’);cat=b.dataset.c;
document.querySelectorAll(‘.uc’).forEach(function(t){var u=U[+t.dataset.i];t.classList.toggle(‘dim’,cat!==’All’&&u[0]!==cat)});
})});
document.querySelectorAll(‘.uc’).forEach(function(t){t.addEventListener(‘click’,function(){
document.querySelectorAll(‘.uc’).forEach(function(x){x.classList.remove(‘sel’)});t.classList.add(‘sel’);
var u=U[+t.dataset.i];document.getElementById(‘dh’).textContent=String(+t.dataset.i+1).padStart(2,’0′)+’ · ‘+u[1]+’ · primitive: ‘+u[2];
document.getElementById(‘dt’).textContent=u[3];setTimeout(post,60);
})});
})();