A failed experiment made the boundary clearer
The historical structured-output decision describes recurring handoff-format regressions. A readable response could still fail the receiving system’s contract. The response was to make the handoff structure explicit and validate it.
A separate implementation review found that successful-looking output did not prove every intended worker ran. Those are different failure modes, but they share a lesson: an assertion about execution needs evidence appropriate to the action.
I use deterministic mechanics where a rule can be checked repeatedly, and model judgment where interpretation and synthesis are needed. Human direction and review remain part of the work. The experiment explanation identifies what the records support and what they do not.
Presentation notes
Read the chapter for the decision, evidence and its limits.