Applied AI Engineering

Structured Outputs & Validation

Structured output turns model responses into data your program can consume. The schema is the start; validation, authorization, business rules, and fallback behavior make it production-safe.

A schema is an API contract
Syntactic validity is not semantic correctness
Unknown must be representable
Validation belongs outside the model
Repair retries need limits

Treat model output as untrusted input

raw = model.extract(document)
parsed = InvoiceSchema.parse(raw)       // shape + type validation
require_user_can_access(user, parsed.accountId)
require(parsed.amount >= 0)
require(evidence_supports(parsed, document))
save(parsed)

The schema gets you parseable data; it does not prove the data is true or allowed.

Validation layers
Layer	Example
Syntax	Valid JSON
Schema	Required fields, enums, numeric types
Business rule	Amount cannot be negative
Authorization	User may access accountId
Evidence	Extracted value appears in source
Safety	Action is allowed without human approval