The model answers in types.
Jev is TypeSafe's System One model. You hand it a state object and a set of typed questions. It hands back a choice, a score or a probability, with the full distribution over the options. There is no prose to parse and nothing to regex.
It reads literally and it is not a calculator, so every number in the state is computed in code and sent as a value and a named bucket. Token names from the chain are truncated and marked untrusted. The state stays under sixty fields.
Confidence policy
| Verdict | Condition |
| killed | WARDEN returned a kill flag. Jev's answer is recorded and ignored. |
| buy | Jev chose buy, confidence is at least 0.55 and rug score is below 1.0. |
| copy | Jev chose copy and at least one tracked wallet is already in. |
| skip | Any buy or copy that misses its condition, downgraded with the reason stated. |
| pass | Jev chose pass. The launch is not worth watching. |
The four questions, exactly as the API sends them
{
"action": {
"type": "choice",
"instructions": "What should a sniper do with this launch right now?",
"criteria": {
"buy": "Enter now",
"copy": "Enter only because tracked wallets are in",
"skip": "Not now, keep watching",
"pass": "Never, walk away"
}
},
"rug": {
"type": "score",
"instructions": "How likely is this token to be a rug or a honeypot?",
"criteria": ["Clean", "Some flags", "Likely rug"]
},
"momentum": {
"type": "boolean",
"instructions": "Buy flow over the last blocks is accelerating."
},
"crowd": {
"type": "boolean",
"instructions": "At least one tracked wallet already bought this token."
}
}
No confidence field comes back. Confidence is the highest probability in the returned distribution, computed in code.