Back to the brain

Docs

JEV is a System One brain. You give it state; it gives you decisions as data. No chat, no generated text, every answer timed.

What JEV is

Agents waste their most expensive calls on forks: is this spam, which worker is next, is this ready for a human. Those are decisions, not writing. JEV is the cheap brain that makes them, so the writer only writes.

The split

creates→ the LLM: research, drafts, explanations.
picks · scores · gates→ JEV: typed questions with explicit options.
does→ code: limits, retries, saved progress, human gates on anything irreversible.

Question types

ChoicePick exactly one labelled option. Returns choice + confidence.
ScoreAn ordered label (low → critical). Returns label + confidence.
NoulYes or no. Returns answer (boolean) + confidence.

Confidence is how sure the brain is about its own answer, not an accuracy guarantee. Gate on it: act above your threshold, escalate below it.

POST /api/decide

Send text as state. JEV answers four typed questions about it: topic (Choice), is_crypto (Noul), urgency (Score) and narrative (Choice).

# request
curl -X POST https://this-site/api/decide \
  -H 'content-type: application/json' \
  -d '{"text":"The new AMM pool lets LPs provide liquidity in concentrated ranges and earn swap fees on every trade."}'

# a real response (fields trimmed)
{
  "decisions": [
    { "type": "choice", "q": "topic", "choice": "defi", "confidence": 0.87, "us": 86.7 },
    { "type": "noul", "q": "is_crypto", "answer": true, "confidence": 0.78 },
    { "type": "score", "q": "urgency", "label": "low", "confidence": 0.70 },
    { "type": "choice", "q": "narrative", "choice": "crypto", "confidence": 0.97 }
  ],
  "us": 187.8
}

Rate limit: 60 calls a minute per IP. Free while it's in this form.

GET /api/state

Counters, latency (median and p99, in microseconds), narrative share, decisions per minute for the last hour, model facts and the latest decisions.

GET /api/stream

Server-sent events. One event per decided launch, the moment it happens:

const es = new EventSource('/api/stream');
es.onmessage = m => console.log(JSON.parse(m.data));
// { kind: "create", name, symbol, mint, decisions: [...4 typed answers], us }

The cortex

JEV subscribes to pump.fun's public feed of new tokens. For each launch it reads the token's own metadata and answers:

narrativeChoice · ai / animal / politics / meme / crypto / other
metadataScore · bare / partial / complete (socials, site, description)
copycatNoul · does the name or ticker copy a launch from the last 24 hours?
dev_buyScore · none / small / large / whale, from the creator's opening buy
These are descriptions, not signals. JEV doesn't tell anyone to buy or sell anything.

The model

The topic head is a multinomial naive Bayes model, retrained on every boot from about 1,000 real crypto pages (docs, research, governance, security, specs) gathered by headless browsers. Accuracy is measured on a held-out 20% of pages it never trained on, and shown live on the home page. The narrative, metadata, copycat and dev-buy heads are lexicon and rule based. Everything runs on one CPU, in microseconds.

Limits & honesty