In a report, the checking mostly goes to the findings. Are the figures right, are the sources in order, is the analysis sound? That's where the effort goes. The recommendation gets a read-through, and that read-through asks whether it sounds reasonable.
Sounding reasonable isn't the same as being carried. A recommendation isn't a summary of the findings. It's a step from this is the case to do this, and that step always contains something no finding says: an assumption about what works, about who can carry it out, about what it costs.
That's not a flaw in itself. A recommendation is supposed to be a judgment. It becomes a flaw the moment the judgment presents itself as a finding.
Why AI sharpens this
A language model writes that step fluently. Three findings roll out into a recommendation that looks logical, with no seam.
There's research on how models handle scope, although it covered summaries and not recommendations. Uwe Peters and Benjamin Chin-Yee compared 4,900 summaries of scientific texts, produced by ten widely used language models, with the originals. Most models stated conclusions more broadly than the source, even when explicitly asked to be accurate. A recommendation resembles a conclusion with a verb attached. The study wasn't done on that, so don't read it as proof. As a reason to look harder, it's enough.
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 an Edit problem. The question for that phase is: does the source still cover the text after editing? For a recommendation, the findings are the source. The text is the recommendation as it ends up in the executive summary, the cover letter or the closing remarks.
That last part is easily forgotten. A recommendation that hangs neatly on two findings in the report shifts in the executive summary, gets shortened in the letter to the council and is a single sentence in the slides. Each step is an edit, and each edit can make the recommendation broader or more definite.
The support table
Make a table with four columns and put every recommendation in it.
- Recommendation
- Carried by
- Scope of those findings
- The leap
Carried by are the findings the recommendation rests on, by number. If you can't point to any, it's an opinion. That's allowed, as long as it says so.
Scope is what those findings cover: which group, which place, which period.
The leap is what the recommendation says beyond the findings. Write it in one sentence. If you can't, that's the reason the recommendation can't be checked.
A made-up example:
- Finding 1: "Of the twelve councils we spoke to, nine say the intake takes longest."
- Finding 2: "In two files a resident was asked the same question three times."
- Recommendation: "Introduce a single digital intake in all councils."
- Scope: the twelve councils we spoke to, not all of them.
- The leap: that a digital intake shortens the process, and that this holds for every council.
The findings say something about what's slow. The recommendation says something about what helps. That's a different kind of claim, and without a second source or a judgment stated alongside it, it isn't covered.
Three tests
1. Strike the finding. Remove finding 1 and read the recommendation again. Is it still standing? Then it rests on something else. That may be expertise or experience, and if so it should say so. In the Deloitte case in Australia, the department said that the substance of the report stood and the recommendations were unchanged, after the invented references had been removed. That may be right. But the question it raises applies to every report: which source carried which recommendation?
2. Could the findings be right and the recommendation still wrong? If the answer is yes, there's a leap. Find it.
3. Who made the leap? A person with the expertise for it, or a draft nobody tested? In TRACES that's the Approval layer: a person confirms the content, and the confirmation is recorded. Fixing a typo needs no expert judgment, rewriting a recommendation does.
What you get for it
A report with a support table is harder to write and much easier to defend. When the client asks "why do you recommend this?", the answer is on the page: these two findings, this scope, and this is our judgment. The judgment is written as a judgment. That's what a client is paying for.
What you can do this week
- Take the latest report and list the recommendations one under the other.
- Next to each one, put the findings that carry it, with their scope.
- For each recommendation, write the leap in one sentence.
- Mark the recommendations you can't tie to a finding, or whose leap is wider than the scope.
- For those, have the author write down where the judgment comes from, and ask who tested it.
The number of recommendations whose leap is bigger than the findings is your Edit leak.
Sources
- Peters, U. & Chin-Yee, B. (2025). Generalization bias in large language model summarization of scientific research. Royal Society Open Science. arxiv.org/abs/2504.00025
- Cyber Daily (2025). Deloitte to refund government after using AI in $440,000 report. cyberdaily.au