Six weeks after a design was agreed, three of its four key decisions are implemented differently and nobody raised it. The document has 18 comments with 16 of them from two people, 11 of 14 engineers opened it once, and the decision record was merged with two approvals in under four minutes. What do you conclude, which signal is misleading, and what would have caught this in week one?

The first three things I would look at

1. The four decisions as written. Are they stated as decisions, with the alternative and the reason, or are they embedded in explanatory prose? If a reader cannot point at the sentence that commits, there was nothing to disagree with.
2. Who the two commenters were. If they were the author and the one other person who already held an opinion, the document recorded a conversation between two people that twelve watched.
3. What the implementers were asked to do after approval. In almost every case the answer is "nothing", and that is the defect.

The diagnosis

Approval was consent to not object, and it was collected in a form that cannot distinguish agreement from absence of objection. Opening a document once is consistent with skimming the summary. A four-minute approval on a decision record is consistent with trusting the author, which is often reasonable and is not the same as having understood the decision well enough to implement it.

The divergence then follows without anyone behaving badly. An engineer reaches decision 2 in code, finds it awkward, picks the option that fits the code in front of them, and does not raise it because from their side nothing has been contradicted: they never held the decision, so they cannot notice departing from it.

The misleading signal

The 18 comments. Comment count reads as engagement and it is the one number here that means almost nothing, because 16 of 18 came from two people. The signal that mattered was the eleven single opens, and it looks like participation. A useful correction is to count distinct commenters, not comments, and to treat a document with fewer than three distinct substantive commenters as unreviewed regardless of its approval state.

The fix

• State each decision as a numbered claim with the option rejected and the reason, so there is a specific sentence to disagree with. Four decisions produce four numbered lines.
• Require restatement, not approval. Before work starts, the engineers who will implement decisions 1 and 2 write one or two sentences each on what they will build and why. This is the read-back used in aviation and surgery, and it works for the same reason: it moves the proof of understanding to the receiver. It costs about ten minutes per person.
• Make divergence cheap to report. A one-line amendment on the decision record beats a re-review. If reporting a change costs a meeting, it will not be reported.

The alert that would have caught it earlier

Not a metric: a checkpoint. The first pull request that touches the area references the decision numbers it implements, and a reviewer checks that. Divergence then surfaces in week one, at a cost of one comment, instead of week six, at the cost of three decisions and the retro.

When this is the wrong answer

For a reversible decision inside one team's own code, this ceremony is waste; the team will find the divergence in review and fix it cheaply. Restatement earns its cost only when the decision crosses a team boundary or is expensive to reverse, which is the same condition that justified writing it down. Skip it entirely when the implementer wrote the design.

The practice also fails in a predictable way: it degrades into a form, and "will implement decision 3 as described" carries no information. Check restatements for paraphrase rather than for existence, and treat a verbatim copy as a non-answer, because the whole value is in the translation.

Common weak answers

• "Add more approvers." More people clicking approve produces more four-minute approvals.
• "They should have read it." True and useless. The process was designed so that not reading it had no consequence until week six.