Decision Primitives
TypeSafe Jev System One
Generative text models fail at sub-second reflex evaluation. Typed semantic judgments provide deterministic gates for agent runtimes.
System One Decision Matrix
System One does not generate streamable prose, code blocks, or conversational filler. It ingests structured state and returns typed machine primitives.
| Primitive | Type Output | Evaluation Contract | Sub-Second Latency | Agent Application |
|---|---|---|---|---|
Choice |
String ID + confidence (0.0..1.0) | Selects exactly one option from a bounded list | 70 to 85ms | Workflow routing, architectural branch selection |
Score |
Float (0.0..10.0) + confidence (0.0..1.0) | Evaluates state against multi-variable criteria | 80 to 95ms | Dependency vetting, code quality smell scoring |
Noul |
Probability (0.0..1.0) | Calibrated probability for a binary assertion | 65 to 80ms | Stop-hook gate, malicious shell command detection |
Safety Calibration
The Confidence Threshold Policy
Text models hallucinate with high conviction. TypeSafe System One returns calibrated confidence scores.
In jev-superpowers, any destructive action, package deletion, git force-push, or branch cleanup requires a minimum confidence of 0.80:
// Rust enforcement in hooks/pre-tool
if action.is_destructive() {
if judgment.confidence < 0.80 {
eprintln!("[BLOCKED] Destructive command rejected: confidence {} < 0.80", judgment.confidence);
std::process::exit(1);
}
}If the oracle returns a confidence of 0.76, the action is blocked by the hook, requiring the human developer to approve the command manually.
Performance Architecture
Speculative Fan-out vs Sequential Prompts
Sequential prompt-chains in coding agents force round-trip network delays of 2 to 5 seconds per question.
System One ingests repository state once and batches independent questions in a single request:
POST https://api.typesafe.ai/v1/systemone
{
"state": "Git staged diff (240 lines) + test failure output",
"questions": [
{ "id": "q1", "type": "choice", "options": ["safe", "risky"] },
{ "id": "q2", "type": "score", "criteria": "Code cleanliness" },
{ "id": "q3", "type": "noul", "assertion": "Tests cover new edge case" }
]
}
// Result: 92ms total for all evaluations · $0.0003 cost