Eval Gates Before You Ship

If you cannot say what got worse, you are not shipping—you are hoping.

The pattern

Keep a small, stable eval suite that blocks merge when critical behaviors regress.

• Golden prompts in the languages you serve
• Tool-contract checks for structured outputs
• Latency and cost budgets as first-class assertions

Minimum viable gate

unit tests  →  golden evals  →  staging smoke  →  release

Start with twenty prompts that represent real customer pain. Expand only when a failure mode repeats.

Trade-offs

• Upfront authoring cost
• Protection against the silent Portuguese regression that English-only suites miss