Tinker Evals & LLM testing

by Kondasamy Jayaraman · drawn with Colorfigures by April Zhu

Preparing the test suite…

Chapter 7 of 7

Release day

The team has a test set, graders it has checked, and a feel for flaky cases. Three new versions wait to replace the one Larkspur ships today. Someone has to decide which, if any, goes out.

This chapter: What rule should decide whether a new version ships, and what can fool it?

You will learn

  • How a release gate turns a test run into ship or do not ship
  • Why a pass-rate bar alone lets a regression through
  • Why the gate is only as good as its grader, and its number of runs
Words you’ll meet (4)
Baseline
The version that ships today. Every candidate is compared against it, question by question.
Candidate
A new version that wants to replace the baseline.
Release gate
The rule that decides, from a test run, whether a candidate ships.
Threshold
The pass rate a candidate must reach to ship, like 75% of the test set.

The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.

Step 1 of 3

Release day. v1 is what Larkspur ships today: the baseline, the version every candidate is compared against. A release gate is the rule that decides whether a candidate ships. This one has a threshold, the pass rate a candidate must reach, and a rule that no question v1 passed may fail. It starts at 50% with that rule off, so v2 ships, though chapter 1 showed it breaks a question. Set the gate so it ships v3 and blocks v2.

Candidate

v2: Bigger pieces: 48 words. Baseline v1: 32-word pieces, the best 3 sent (RAG’s defaults).

Grader