Topics in the media

Trust fell because people looked closely; unreviewed merges rose because looking closely does not scale. Both are rational and they point in opposite directions. Adding reviewers treats a structural property as a staffing gap: review needs context you cannot hire quickly, parallelises badly, and degrades silently rather than loudly.

The waterfall comparison is lazy. Waterfall failed because the specification was written once, by people distant from the work, and could not respond to what building taught them. A criterion attached to a ticket, revised in refinement and checked on every change, is the opposite of that. The real question is not whether to specify but who does it and whether a machine can settle it.

Generated code is not intrinsically unmaintainable. Unread code is. The distinct category is comprehension debt: a system being modified faster than anyone reads it, where the cost surfaces during an incident rather than during maintenance. A global workday figure is unactionable; your own unread-merge rate is not.

A junior who accepts output they do not understand creates debt at machine speed, which is a training problem rather than a hiring mistake. The pipeline broke because the work juniors learned on automated, and no interview technique fixes an absent apprenticeship. Someone has to rebuild it deliberately.

The provenance gap is the most consequential item on this page. There is an established way to attribute a generated image and no equivalent for a generated function. "Tamper-evident" is also a stronger word than most teams register: a log written by the same pipeline that performed the action is the component attesting to its own behaviour, which is the arrangement controls exist to avoid.