A research report with no gaps should make you suspicious. Real questions almost always have parts that nobody wrote down anywhere the researcher was allowed to look: a figure that was never published, a clause that does not exist, a date on which two documents disagree. The interesting question about a report is not whether it has gaps. It is whether it tells you where they are.
The real failure is the filled gap
When an agent meets a hole in the evidence, it has two options. It can say “I could not find this”, or it can produce something that sounds right. The second is the dangerous one, and not because it is always wrong. A plausible guess is often close. The problem is that a guess written in the same voice as a finding is indistinguishable from one. A reader who cannot tell them apart has to either check everything, which defeats the purpose of the research, or trust everything, which means trusting the guesses too.
The cost is lopsided. An honest blank costs a follow-up question. A confident wrong number in a due diligence memo, a pricing comparison or a regulatory summary can travel a long way before anyone looks at where it came from, and by then other decisions rest on it. Generated text is fluent by design, so fluency proves nothing about support. The control has to come from the structure of the output, not from its tone.
Three kinds of statement, kept apart
It helps to name what a report can contain. Most problems come from mixing these three.
| Kind | What it is | How it should appear |
|---|---|---|
| Finding | A claim supported by a quotation from an allowed source | The claim, the quoted passage, and a pointer to where it sits |
| Open question | Something asked that no allowed source answered | The question, where it was looked for, and the fact that nothing was found |
| Unsupported claim | A statement with no source behind it, often drawn from background knowledge | Flagged as such, or kept out of the findings entirely |
The open question is a result in its own right. “The agreement does not specify a notice period for price changes; checked the signed agreement and both amendments” is useful information. It tells the reader that the search happened, how far it went and where a person should look next.
“Not found” and “unavailable” are different answers
An empty result is a valid answer when the search actually ran across the sources it was meant to cover. It is a different situation when the search could not run: a source was offline, a document failed to import, a permission was missing. Reporting both as “no results” hides the second problem inside the first. A good report says which one happened. “Nothing found in the five approved documents” invites a decision about widening the sources. “Two of the five documents could not be read” invites fixing the import before anyone draws a conclusion.
Scope matters too. “Not found” always means “not found in the sources that were allowed”. That is why the allowlist should be visible next to the open questions; deciding which sources may answer covers how to set it.
How to keep them apart in practice
Most of this can be written into the request before the agent starts. Some instructions that work:
- Every finding includes the quoted passage and a pointer to its location. A finding without a quotation moves to the open questions list.
- No number appears in the summary unless it also appears in a cited passage.
- When a value cannot be found, the cell says “not found” instead of holding an estimate.
- The report ends with an open questions section that lists, for each question, the sources that were checked.
- Failures to read a source are reported separately from searches that returned nothing.
Making the open questions list part of the acceptance criteria is the strongest version: a report without it is not done, however polished it looks.
Reading a report that keeps them apart
Separation helps the reviewer only if they use it. Two habits are worth building. First, read the open questions before the findings. They show where the evidence was thin, which is where a person's time is best spent. Second, treat each citation as a pointer, not as proof. Follow it, read the passage, and check that it says what the finding claims and that it applies to the case in hand. A link the agent generated, or a footnote that looks tidy, is not evidence until someone has read the passage behind it. Citations that hold up goes deeper on what to check.
How GoCux handles gaps
GoCux records a finding backed by a quotation apart from a question that no source answered, and it keeps both. In Extraction, each value arrives with the passage behind it, and a value that is not in the material stays unknown instead of being guessed. FindAll reports the limits of its coverage instead of rounding them away, because a count of results does not prove that a list is complete. The aim in each case is the same: a reader should be able to see the edge of what is known.
What to do with an open question
- Decide whether it matters. Some gaps do not change the decision. Record them and move on.
- Bring the missing material. If the answer is in a document you hold, add it; see bring the evidence you already have.
- Widen the sources with a new plan, approved like the first one, rather than stretching the running work.
- Ask a person. Some answers live in someone's head or inbox, not in any document.
- Accept the unknown explicitly. “We do not know this, and we are proceeding anyway” is a legitimate decision once it is written down.
Bring this workflow into your workspace.
Set the scope. Review the evidence. Approve the steps that need it.