The four checks before anything leaves the building

The last piece in this series priced the verification tax: checking has a real cost, and it is worth paying whenever a task sits outside a tool's known competence or the error would be expensive to miss. That still leaves the practical question unanswered. "Check it" is not an instruction anyone can actually follow. Without a concrete method, checking collapses into one of two failure modes: skimming, which misses everything that matters, or redoing the whole task by hand, which erases whatever time the tool saved in the first place.

This piece names four specific checks that catch most of what actually goes wrong, without requiring anyone to re-derive the entire piece of work from scratch.

Why "check it" is not an instruction

Told to check a document, most people read it once, looking generally competent, and sign off. That reading catches typos and awkward phrasing, and misses almost everything that actually matters: a fabricated figure stated in fluent, confident prose looks exactly like a correct one on a first read. A real check has to target specific, predictable failure points rather than relying on a general impression of quality.

The four checks

• Check the source, not the sentence. Trace every specific number, quote, or named fact back to where it actually came from, because a wrong figure never announces itself; it sits in a sentence that reads exactly like every correct one around it.
• Check the boundary case, not the typical case. Most of a document is the easy, well-covered middle; errors concentrate at the edges, in the exception, the smallest or largest figure, the case that almost didn't fit the category it was put in.
• Check the judgement call, not the fact. Somewhere in most delegated tasks a genuine judgement was made, a risk flagged or waved through, a recommendation chosen over an alternative, and that specific call deserves scrutiny that a routine fact does not.
• Check what is missing, not just what is there. An omission is invisible on the page, which is exactly why it survives a normal read; verifying completeness against the original request catches what a fluent read never will.

A check that only reads what is there will never catch what was left out.

The trade-off in a four-check method

Four targeted checks are not exhaustive, and an error sitting outside all four categories can still slip through; a genuinely thorough audit would touch everything, which is precisely the redo this method is built to avoid. The honest answer is that the four checks are a deliberate trade: they concentrate limited checking time on the highest-yield failure points rather than spreading it evenly across a document, on the basis that most real errors cluster in sources, edges, judgement calls, and omissions rather than in the well-covered middle.

The practical takeaway

Before anything produced with AI assistance leaves the building, run these four checks: trace the sources, inspect the boundary cases, scrutinise the judgement calls, and verify against the original request for what might be missing, rather than relying on a general read-through.

No external statistic cited; this article presents an internal verification checklist rather than third-party evidence.