Having everything checked sounds sound. In practice, a list of changes gets skimmed. A draft report easily has forty edits, nearly all harmless: a comma, a shorter sentence, a different word for "implement". Read them all with equal care and you read the four that matter as lightly as the rest. Read them all lightly and you miss those four.
So the answer isn't more checking. It's deciding in advance which changes need a judgement, so attention goes where it's needed.
Which phase this is
SEAL describes the four phases knowledge work goes through: Source, Edit, Approve, Launch. In every phase the traceability of a source can be lost, and the weakest phase determines what the work is worth.
This is Approve. The question for that phase is: who has assessed the content, with what expertise, and on what basis?
Approve usually fails because nobody has written down what has to be assessed. Then you get one of two outcomes, and neither works. Either everything goes past a subject-matter judgement, and the oversight quietly evaporates in a busy week. Or nothing does, because "it's only a small tweak". The phase works the other way round: not more eyes, but fewer moments. Only the risky changes need a judgement, and it has to be a real one, with the source alongside and the outcome recorded.
The cause: size isn't the risk
It's tempting to sort changes by size. Rewriting a paragraph feels heavier than swapping a word. But the risk lies in what a change does to the claim.
The smallest word there is, is "not". Shifting one figure, turning "most" into "all", or "can" into "will": single-word changes that make the text claim something different. Rewriting a whole paragraph to improve the flow may change nothing of substance.
So the dividing line is this: a change needs a judgement if it alters what the text claims, not if it alters how the text sounds.
The test: three bins
Go down the list of changes and drop each into one of three bins.
1. Read along. Spelling, layout, punctuation, paragraphs reordered without shifting the meaning. One glance is enough. No tick needed.
2. Check against the source. A sentence that makes a claim has been reworded, but it should still say the same. Whoever makes or reviews the change puts the sentence next to the source passage and sees whether it still says the same thing. That takes a minute and doesn't need a specialist.
3. Judge. The change alters what is claimed. Think of:
- a figure, name, date or amount;
- a quantity or scope ("some of", "most", "all");
- the level of certainty ("can", "seems", "will");
- a conclusion or recommendation;
- a legal or financial passage;
- anything an AI system wrote or rewrote without a source attached.
For a judgement, someone with the subject expertise decides, with the source next to the sentence, and the outcome is recorded. The approval belongs to the passage: if it's changed again later, the approval lapses. That's the core of Edit, and it's why Approve only works when the two phases connect.
If you're unsure which bin, pick the heavier one. It costs a minute now and then. The other way round costs you an error under your name.
An example
A made-up example, to see the bins at work. A draft advice paper has fourteen changes:
- Nine are read-along: three typos, two headings, four commas.
- Three go to the source: sentences shortened, where the source passage shows nothing has dropped out.
- Two need a judgement. In the first, "about six weeks" became "six weeks", which removes the margin from the schedule. In the second, "is recommended" became "is essential", which turns advice into a requirement.
That's two judgements out of fourteen, and that's the point. Someone with expertise spends a quarter of an hour on what needs expertise, and the rest of the list takes five minutes.
What to ask of a system
If you use a tool to write or edit, the same question returns in TRACES, at the Approval layer: does a person confirm the content, differentiated by the kind of change, and is that confirmation recorded?
TRACES are the six layers on which you assess an AI system: Trail, Retrieval, Approval, Constraints, Expertise, Storage. Where a layer is missing, there is a risk.
A system that treats every change the same doesn't do it. Asking for approval on everything leads to clicking through. Asking for nothing stops being oversight. A usable system shows which bin each change falls in, displays the source passage for a judgement, and keeps a record of who approved what and when.
What you can do this week
- Take the last document you revised, ideally with tracked changes or an earlier version alongside. One that was rewritten with AI works too.
- Sort each change into one of the three bins. Count them.
- For each bin, look at what actually happened. Did someone review the bin 3 changes with the source alongside, or were they read along?
- Write one page on what needs a judgement in your organisation: five to seven lines, in your own words, based on the list above. Attach it to the workflow, not to a folder nobody opens.
The gap between bin 3 and what was actually checked is your leak at Approve.