From pilot to reliable flow: schema, sources, abstention and validation

What you will achieve now

A decision or test applied to "From pilot to reliable flow: schema, sources, abstention and validation".

• 1 Understand the criterion
• 2 Do a small practice
• 3 Save evidence

A short note with what you did, what went well, what failed, and what you would review next.

• Turn a vague request into a contract of input, output and limits.
• Make a document-based flow cite its evidence or abstain.
• Validate critical fields before a person approves the result.

The difference between a demo and a flow

A demo gives an apparently good answer to one example. A flow must respond in a controlled way to many examples, including incomplete and wrong ones. To achieve that, separate the stages: receive an authorized input, clean what is unnecessary, retrieve evidence if it works with documents, generate a structured draft, validate simple rules and hand it off for review. Only a later, explicit action could use that result.

Entrada autorizada
  -> normalizar y minimizar datos
  -> recuperar fuentes permitidas si hacen falta
  -> generar borrador con esquema fijo
  -> validar formato, campos y reglas de negocio
  -> revisión humana
  -> acción posterior solo si se aprueba

Write the contract before the prompt

The prompt helps guide the model, but it is not the full contract. First define the work without mentioning brands or models. For extracting an invoice, for example, the output is not "a useful summary": it is a draft with concrete fields, an identifiable source and a list of open questions. If a field is not found or fails a basic rule, it should stay empty and marked for review.

nombre: extracción de factura a borrador
finalidad: preparar una fila revisable; no contabilizar
entrada: PDF autorizado y su identificador interno
datos_no_permitidos: contraseñas, datos ajenos al documento
salida: emisor, fecha, número, base, IVA, total, dudas, fuente
regla_de_evidencia: cada campo crítico indica página o zona leída
abstención: si no puede leer o no encuentra un campo, usa null y explica por qué
validación: fecha válida; importes numéricos; total coherente o marcado
revisor: administración
acción_prohibida: exportar a contabilidad, pagar o modificar registros

Sources: evidence before apparent certainty

If the flow answers about a policy, a catalog or a document base, the model must not fill gaps with plausible phrases. Deliver alongside the answer the identifier, version and fragment of the source that supports it. If there is not enough evidence, the correct output is a useful abstention: "I cannot find an authorized source to claim this; review this document or escalate the query".

• Limit the search to the corresponding set of documents and permissions.
• Store version, date and fragment of the source, not only the final text.
• Distinguish an extracted fact, an inference and a drafting proposal.
• Do not turn a similarity score into a guarantee that the answer is correct.

Validate outside the model what is objective

A model can propose a date, an amount or a code, but a deterministic rule must check afterwards that the format exists and makes sense. Validate types, required fields, ranges, duplicates and simple relationships. Human review decides meaning; rules stop an incomplete output from looking ready to use.

salida_borrador:
  emisor: ACME S.L.
  fecha: 2026-07-18
  numero: F-204
  base: 100
  iva: 21
  total: 121
  evidencia:
    total: pagina 1, bloque inferior
  dudas: []

validaciones:
  - todos los campos críticos existen o están marcados null
  - total = base + iva, con tolerancia explícita
  - fecha se puede interpretar
  - número no está duplicado en la muestra
  - si falla una regla: estado = revisar, nunca = listo

If you have saved the evidence from this lesson, continue with "Failure tests and metrics". If not, repeat the check before moving on.

Learn from a verifiable reference

Aulafy separates stable concepts from data that change —versions, prices, models and commands—. Check the course sheet for the review date, the verified scope and the primary sources.

"Editorial review" means structure, claims and sources have been reviewed. It does not mean every command has been executed: when a technical test exists, it will be labeled as such.