gabrielwang.ai ← back to the workshop
Edition 013 · interactive demo

The suite is green. The code it tested is gone.

If you are commissioning AI-built systems, the question you actually need answered is "how would I know if it stopped working?" — and the usual answer, "there are tests and they pass", is weaker than it sounds. Here are five ways a suite stays green while the thing it verifies is broken. All five are real; all five were found in one week, in one codebase, by the person who wrote them. Pick a case and undo the fix yourself.

Honest-AI note. No model runs on this page — and it isn't really a page about AI. It is about the discipline that makes AI-written code safe to depend on, because an agent that writes its own tests can write five green tests that prove nothing just as easily as a human can. Every result below is computed live: each case is a working pair of implementations and a working pair of locks, and the PASS/FAIL cells come from running them in your browser, not from a table someone typed.

Synthetic fixtures Locks really execute Runs in your browser No live model call

The lock as written

          
The honest lock

          

Plain-language key (lock, fixture, gate, vacuous pass, bytecode cache)
Lock
A test written specifically to keep a past correction from being undone. Its whole value is that it turns red when the correction disappears.
Fixture
The sample data a test runs against. When the code changes shape and the fixture doesn't, the test can quietly stop touching the real path.
Gate
The check that must pass before work is released. A gate you only run when you remember is not a gate — wire it into the commit and prove it blocks.
Vacuous pass
"All of them are fine" is automatically true when there are none. Any check of that shape must first prove it found something.
Bytecode cache
A compiled copy the interpreter keeps to start faster. If it isn't invalidated, the suite can run code you have already edited away.