Behaviour people can depend on

I define acceptance criteria, test varied and ambiguous inputs, and investigate inconsistent outputs. The aim is to make unreliable behaviour easier to find and give the team useful evidence for deciding where a feature needs tighter limits or human review.

Questions to resolve before evaluation

What user task is the feature meant to support?
What makes an answer or action acceptable for that task?
Which variations, ambiguous requests, and failure cases matter?
When should the product ask for clarification, decline, or hand control to a person?
How will a change in behaviour be detected and investigated?

Evidence for better decisions

Evaluation scenarios, reviewed outputs, and documented inconsistencies help the team see which tasks the feature handles well and where it falls short. I make the evaluation conditions and gaps clear so a successful sample is not mistaken for dependable behaviour across every use case.

Connected to software quality

An AI response can look reasonable while violating a workflow rule. The surrounding interface, permissions, API behaviour, error handling, and confirmation steps still need testing. My QA and test automation work and product engineering background help connect those concerns.

Read the practical analysis

What Gemini’s image-generation failures teach us about AI acceptance criteria examines a public incident and proposes a task-based evaluation approach.