What happened when we teased the semantics out of the scenarios.

I remember looking at the test suite the team had put together - around 500 integration tests covering our API surface area. Alongside the usual happy-path and negative cases, we even had a few that invoked the APIs concurrently: create two items at once and assert that only one call should succeed, or create and delete an item at the same time and assert that they should happen in some logical order (the item should never be in half created/deleted state). Those concurrent ones were a mind bender, because what you observed depended on the logical order the operations ran in. A developer had to simulate the orderings in their head and write the assertions for each. Overall, we were reasonably proud of our test suite.

Yet I still had a nagging feeling.