← Glossary

Evaluation-Driven Development

A methodology for building AI features where each feature is treated as a testable hypothesis and the pull request is gated on an offline evaluation pipeline -- gold-standard and synthetic datasets scored by code-metric and LLM-as-judge evaluators -- instead of on traditional unit tests.

Context

Evaluation-Driven Development (EDD) was the deep dive on Episode 31 of the ADI Pod, drawn from a Decoding AI piece by Paul Easton and Alejandro Aboy. The premise: a non-deterministic AI feature can’t be gated by a unit test the way deterministic code can, because the same input can produce a different output run to run. EDD replaces the unit-test gate with an offline eval pipeline — in the article’s case built on Opik — and treats every feature as a hypothesis you measure rather than a spec you assert.

The moving parts the episode pulled out:

Why It Matters

Shimin’s framing on the show was that moving from unit tests to evals is like moving from Newtonian physics to quantum mechanics: the old rules don’t vanish, they stop being the right level of description once the system is probabilistic by design. EDD is the non-deterministic era’s analog to test-driven development — you build the measurement in before the feature, and you let the eval, not a green checkmark on a unit test, decide whether the PR ships.

It also fits the show’s recurring throughline that the leverage in AI engineering keeps moving up the stack: the hard part is no longer writing the feature but specifying how you’ll know it works, which is exactly the judgment a human still has to supply.