JEV Intelligence
Jev is an external AI model, TypeSafe's System One, that reads each new Robinhood Chain launch's name, ticker, description and links and returns calibrated probabilities that become five filterable token parameters.
In Neuron code, "Jev" is specific: it is TypeSafe's System One model, pinned to version jev-1.13.0, called by the indexer (indexer/jev.ts). It answers questions about the parts of a launch that numbers cannot express, such as whether a name is a copycat or a description is pure hype. It does not see prices or trading, does not decide trades, and does not run as a separate pre-trade check. Its answers are stored as token parameters, so an agent uses Jev only when its filter says so.
What Jev reads
The indexer turns each launch into plain-language facts first. Jev is weak at arithmetic, so counts are bucketed into sentences in code:
| Input | What is sent |
|---|---|
| Name and ticker | As declared by the token, or "(none)" |
| Description | The first 800 characters, or "(no description)" |
| Links | X, Telegram, website, Discord and Farcaster, or "(not set)" for each |
| Ticker history | "This is the first token with this ticker", or how many earlier tokens used it |
| X link history | Whether earlier tokens already used the same X link, and how many |
| Deployer history | First launch, or how many earlier launches and how many graduated; over 20 earlier launches is called out as a serial launcher |
What Jev returns
Five questions are asked in one call. Each answer is stored on the token and exposed as a filter parameter:
| Parameter | Question | Output |
|---|---|---|
jev_rug_prob | Does the launch look low-effort or deceptive (copied or generic name, placeholder or mismatched links, recycled ticker or X link, serial launcher, text promising profits)? | Probability, 0 to 100 |
jev_copycat_prob | Does the name or ticker imitate a well-known token, brand, celebrity or trending coin? | Probability, 0 to 100 |
jev_hype_prob | Does the description rely on hype or urgency (moon, 100x, guaranteed) instead of saying what the token is? | Probability, 0 to 100 |
jev_effort | How much real effort do the name, description and links show? | Expected level from 0 (none) to 3 (detailed project with consistent links) |
jev_narrative | Which theme is the token about? | One of ai, political, celebrity, animal, finance, gaming, culture, utility, other |
jev_effort is the probability-weighted average of the four levels, not a single pick, so 1.6 means "most likely between some and minimal effort". The model also returns a confidence, which is stored but not offered as a filter.
When tokens are scored
- Only launches whose metadata has been read, launched within the last 2 days.
- Newest first, with tokens trading in the last hour ahead of the rest.
- Batches of 24, six requests in parallel, polling every few seconds; scores usually land within minutes of launch.
- A token that fails three times is skipped. Until scored, the
jev_*values are empty, and an empty value never passes a filter. - Scoring runs only while the indexer has a TypeSafe key configured.
The model version is pinned on purpose: a model update cannot silently shift the probabilities that rules were tuned on.
How it feeds decisions
Jev has no say unless a rule uses it:
- 1As a filter. Add a condition such as
jev_rug_prob < 70orjev_narrative is aiin the Custom Strategy Builder. None of the prebuilt templates use Jev by default. - 2In the preview. The "Matching right now" table on /create shows each token's Jev rug probability in a "Rug" column, so you can see what a Jev condition would remove.
- 3As context for the AI pre-trade check. When that guard tool is connected, Claude is given the token's full indexer record, which includes the Jev values, before it decides whether to veto.
- 4In backtests. Jev values are recorded in the 5-minute snapshots like every other parameter, so a Jev condition can be backtested once history exists.
Is it working?
The indexer measures Jev against what actually happened and publishes it on the Lab page (/lab). It takes launches 6 to 48 hours old (old enough to have played out) and groups them by jev_rug_prob band (0 to 50, 50 to 70, 70 to 80, 80 to 90, 90 and up), both for every launch and for launches with at least 50 makers, and by jev_effort band. For each band it reports:
- how many graduated to a Uniswap pool;
- the median peak multiple (
ath_x); - the share that reached at least 2x their launch value;
- the share that dumped, giving back 75% or more of their peak.
Jev judges how a launch presents itself, not whether it is safe. A polished launch can still be a rug, and a silly one can still run. Treat a Jev condition as one filter among several, and check its value on /lab before relying on it.
Related
Neuron is on Robinhood Chain testnet. Testnet performance is not evidence of real-money profitability.