Playground: siente lo que hace un modelo de juicio
Con la misma entrada: a la izquierda, la salida típica de un modelo generativo; a la derecha, el juicio tipado de Jev. Todas las respuestas son resultados reales de la API pregrabados (jev-1.13.0) con el uso real de tokens. Edita la entrada — la respuesta no cambia: pertenece a la entrada original grabada.
Primitiva · Noul
¿Es urgente este mensaje de soporte?
Noul evalúa una pregunta de sí/no y devuelve la probabilidad de «sí»: un valor de 0 a 1.
I'm really sorry to hear you've been having trouble connecting your Stripe account. I understand how frustrating that must be, especially when it affects your sales. Let me help you look into this. Could you tell me a bit more about...
Choice elige una opción de un conjunto fijo, con probabilidades por opción y una confianza general.
questions (la solicitud)
{
"intent": {
"type": "choice",
"instructions": "What is the primary intent of this message?",
"criteria": {
"bug_report": "Reporting something broken",
"feature_request": "Asking for new functionality",
"how_to_question": "Asking how to do something",
"other": "None of the above"
}
}
}
Un modelo generativo produciría
Thanks for reaching out! It sounds like you are experiencing an error with the export functionality on the reports page when using Safari. This could be related to browser compatibility. Have you tried...
Score califica según criterios ordenados; el valor puede quedar entre niveles, con probabilidades por nivel y confianza.
questions (la solicitud)
{
"severity": {
"type": "score",
"instructions": "How severe is the issue?",
"criteria": [
"Minor",
"Moderate",
"Critical"
]
}
}
Un modelo generativo produciría
I'm so sorry to hear that your payouts have been failing for three days. That sounds really stressful. Payout failures can have several causes, such as... (510 more words)
Jev devuelvePregrabado · no es inferencia en vivo
score = 1,87 (entre Moderado y Crítico) · confianza 0,81
Combina dos puntuaciones independientes —intensidad emocional e impacto de negocio— en una prioridad de respuesta ponderada. Tipado y explicable.
questions (la solicitud)
{
"emotional_intensity": {
"type": "score",
"instructions": "How emotionally intense is this message?",
"criteria": [
"Calm and factual",
"Frustrated but professional",
"Angry or escalating"
]
},
"business_impact": {
"type": "score",
"instructions": "How severe is the business impact described?",
"criteria": [
"No material impact",
"Degraded operations",
"Revenue-blocking outage"
]
}
}
Un modelo generativo produciría
Dear valued customer, thank you for bringing this to our attention. We sincerely apologize for the inconvenience caused by the API errors. Our engineering team is investigating the issue with the highest priority...
Sin «clasificar y luego preguntar». Envía de una vez todas las preguntas que puedas necesitar — las respuestas son independientes y tu código decide cuáles usar. Los tokens de salida son gratis.
questions (la solicitud)
{
"request_type": {
"type": "choice",
"instructions": "What kind of request is this?",
"criteria": {
"sales_question": "Asking about plans, pricing, compliance, or contracts",
"billing_issue": "Payment, invoice, or refund matters",
"technical_support": "Something is broken or needs troubleshooting"
}
},
"is_existing_customer": {
"type": "noul",
"instructions": "Does the writer appear to be an existing customer?"
},
"tone_politeness": {
"type": "score",
"instructions": "How courteous is the tone overall?",
"criteria": [
"Blunt or rude",
"Neutral",
"Warm and appreciative"
]
}
}
Un modelo generativo produciría
Thank you for reaching out! To answer your questions: regarding SOC 2 report access, our enterprise plan does include... (the answer continues for 300+ words, mixing both topics with no typed structure to route on)
Jev devuelvePregrabado · no es inferencia en vivo
choice=sales_question (conf 0.25) · noul=0,86 · score=1,74 — una llamada, tres decisiones independientes
JevCode