The Evaluation Loop Is Part of the Product

An evaluation spreadsheet is useful. An evaluation loop is better.

The static-test trap

A team creates a golden dataset, runs the model and reports accuracy. The number moves, but nobody can explain which architectural change produced the movement.

That is testing without diagnosis.

A stronger loop

production failure
      ↓
reproducible case
      ↓
failure taxonomy
      ↓
architecture / prompt / data change
      ↓
offline evaluation
      ↓
trace review
      ↓
release

Each new failure should either fit an existing category or force the taxonomy to become better.

Metrics need evidence

Aggregate accuracy can hide regressions in a high-value slice. Keep enough metadata to answer which intent, data source, language, tool path or reasoning pattern failed.

The test suite becomes product memory

A mature evaluation set is a record of behaviors the product has promised not to forget.

That is why evaluation infrastructure should evolve alongside the application rather than arrive as a final QA step.