Evaluation and Quality

Evaluation should not produce an abstract score. It should decide whether the output meets release-ready acceptance criteria.

Evaluation layers

Layer	Question	Evidence
Single case	Did one input produce the right output?	Fixture, snapshot, assertion
Scenario	Did a workflow complete?	Harness scenario, trace
Regression	Did old behavior break?	Test suite, link audit
Release	Is this ready for public users?	Build, accessibility, content review

Evaluator design

A good evaluator:

• Checks one clear standard.
• Explains failure reasons.
• Handles boundary input.
• Matches human acceptance criteria.
• Stays stable across model or prompt changes.

Documentation quality gates

This project should run at least:

npm run docs:build
npm run docs:check-links

Manual checks:

• Homepage first screen is clear.
• Mobile layout does not overflow.
• Chinese and English navigation is complete.
• New pages include concepts, steps, practice, troubleshooting, and next links.

Practice

Write 5 acceptance criteria for a new MCP tutorial page and turn 2 of them into automated checks.

Next: Deployment and Safety.