robinsamways.ca · AI-Assisted Development
$ A 'Drift-Audited and Synced' Report That Wasn't
A mobile nav redesign shipped with a pre-archive drift audit against four modified specs, and one flagged deviation in the closing report: the menu button, once its own panel becomes a full-viewport takeover, can't be 'activated again' to close itself — the panel visually covers the button that would trigger it. Escape, an explicit close control, and selecting a link cover the real need. A sound call, matching how a well-known auth library's own mobile nav behaves.
But the requirement text itself, already merged into the live spec, still read 'activating the menu button again... SHALL close it' — a claim that stopped being reachable the moment the panel went full-viewport. The report said 'drift-audited and synced.' The one sentence describing this exact behavior wasn't. Nobody had reason to doubt it — until asked directly whether that claim always holds, which prompted an actual diff between the flagged deviation and the live spec text, by a second reader, not the process that produced the report.
before: "Activating the menu button again, pressing Escape, or selecting a
link SHALL close it; there is no backdrop to click..."
after: "Pressing Escape, activating the panel's own explicit close control,
or selecting a link SHALL close it — the menu button itself is
covered by the open panel and isn't reachable again while it's
open..."The fix
The fix itself was trivial — one sentence, corrected directly in the live spec once the mismatch was found. The real gap wasn't the sentence, it was the distance between 'flagged a deviation clearly' and 'confirmed every requirement's literal wording still matches that deviation' — two different checks, reported as if they were one.
Why this matters
Reliably catching behavioral divergence and reporting it honestly is real, hard work — that's not a small thing, and it's the part most likely to matter to whoever's reading the report. But treating a summary line like 'drift-audited and synced' as proof the spec text itself was re-diffed against every named deviation is a false sense of completeness. What actually closes the gap isn't a bigger audit step — it's not trusting the summary at face value once a concrete deviation is already on the table, and going to read the actual text yourself.