How do you upgrade a large production system that has almost no tests?
So far: Start by writing down what is not being done, map the seams, and pin behavior with characterization tests before extracting anything.
Read the write-up →Innovation lab · Research
Questions worth answering.
Research here means a clear question, evidence gathered carefully, the smallest test that could prove me wrong, and a written answer.
How I research
- Frame the question. One sentence, and what a useful answer would change.
- Gather evidence. Read the code, instrument the system, count things. Avoid opinions that cannot be checked.
- Test the smallest thing. A characterization test, a proof of concept, a side-by-side comparison.
- Write it down. A dated decision with the alternatives and the tradeoff, so the reasoning outlives the meeting.
Current and past questions
Ongoing
Answered
Answered
Ongoing
Exploring
Answered
What I am working on.
What should identify a record when two systems both think they own it?
So far: One owner for identity, an id we control, and never sending an important field blank.
Read the write-up →How do you make a very large data move safe to repeat and easy to prove?
So far: Append-only audit log, skip what already succeeded, and finish with a read-only check of what landed.
Read the write-up →How much of a permissions model can be generated instead of hand-tested?
So far: More than two thousand role checks came from one matrix. The open question is how to review the matrix itself.
Read the write-up →How can AI assistants speed up reading and planning without lowering the bar?
So far: Treat their output like a pull request from a new teammate: reviewed, tested, owned. Still refining the checklist.
Read the write-up →What does code-first look like for a whole company setup?
So far: CRM, hosting, DNS, brand and sign-in all defined in reviewed scripts with a verify step.
Read the write-up →
Methods
The toolkit.
Seam mapsCharacterization testsDated decision logsRisk registersProofs of conceptSide-by-side comparison harnessesGenerated testsResumable runs with audit logsRead-only verificationStatus documentsHandoff notes