Architecture
One Shot at Understanding
The model gets one shot at understanding the question. The person gets unlimited cheap corrections. Almost every other property of the design falls out of that asymmetry.
Writing
Engineering craft — architecture, decisions and what they cost. Writing about the QA category itself lives on the Hawzu blog; those pieces are listed here too, marked ↗.
Architecture
The model gets one shot at understanding the question. The person gets unlimited cheap corrections. Almost every other property of the design falls out of that asymmetry.
blog.hawzu.com
Building a test case generator turned up rules the model ignored, and one that never reached it at all. What moved into the schema, what could only be tested, and the two things that came back out.
blog.hawzu.com
The AI is grounded in three separate corpora — the product docs, your own specifications, and your test coverage — kept deliberately apart, each behind exactly one read path.
Architecture
We didn't try to predict every bad request. We made the API surface explicit, then made the application enforce its own definition of what is allowed to exist.
blog.hawzu.com
A blank chart builder gets called flexibility, and it hands the reader a research project. Starting from the question instead is what the presets, categories and role bundles all exist to protect.
blog.hawzu.com
No single metric answers "are we ready to ship?" Release confidence is assembled from connected evidence — coverage, executions, defects, risk — until the pieces point the same way, then judged by people.
blog.hawzu.com
Metrics tell you what you know, never what you don’t. Software is uncertain by nature, so testing can only reduce that uncertainty enough for people to decide — it can’t remove it.
blog.hawzu.com
Modern QA tools keep getting more powerful, but the category still struggles to answer the one question every release meeting asks — are we ready to ship?
blog.hawzu.com
Sign off a release and its analytics stop moving — the defect picture and the acceptance criteria it was judged against are both snapshotted, so a release that shipped as Ready still reads Ready a year later.