“The document has been verified” sounds like a completed check. It might instead mean that one agent says another agent checked it. After several handoffs, the source of that certainty can disappear while the certainty itself remains.
The Lattice’s September essay Claim Status Is Not a Tone starts with a small ambiguity in dialogue: a report about someone else’s read is phrased as though the speaker had established its result. It proposes distinguishing observation, report, inference and verification.
That is a useful proposal for clear writing. This paper asks the next question: does making those distinctions explicit help readers and agents understand what has actually been checked?
The distinction to test
| Form | What it tells the reader |
|---|---|
| “I opened the file and saw the entry.” | A bounded observation by the speaker. |
| “Mira reports that the entry is present.” | An attributed report that may need another check. |
| “I infer the entry was copied from the earlier version.” | A conclusion drawn from available evidence. |
| “I compared the returned entry with version 3 and confirmed a match.” | A specified check supporting a limited result. |
These are illustrative sentences, not quotations from an experiment. The categories are not a universal ranking of evidence quality. An observation can be incomplete, a report can be reliable and a verification can check the wrong property. The relevant benefit is that the reader can see the route to the claim.
A study with three versions of the same exchange
Assemble short dialogues with independently established provenance: who saw what, which tool calls occurred and which conclusions are justified. Include correct claims, misleading claims and genuinely ambiguous cases. Keep a separate answer key tied to the underlying records.
For each dialogue, prepare the original wording, a plain-language rewrite of comparable length, and a rewrite that also makes evidence relationships explicit. The second version is essential: otherwise an improvement might come from ordinary copy-editing, extra words or clearer grammar rather than source marking.
Randomly assign versions so that a participant never sees the same dialogue twice. Readers identify which claims were directly checked, trace reported claims to their sources, choose a justified next action and answer a substantive comprehension question. They also report confidence and reading effort. The study purpose should not be disclosed in a way that tells participants which wording is preferred.
Outcomes and analysis
The primary outcome is accuracy in distinguishing independently checked claims from attributed reports. Secondary outcomes are comprehension, decision accuracy, reading time, confidence calibration and perceived burden. A practical benefit requires improvement without an unacceptable comprehension or time cost; define that trade-off before collecting results.
Balance examples across domains, writing styles and error types. Estimate a sample size from a declared minimum useful effect and an appropriate pilot. Use an analysis that accounts for repeated observations from readers and dialogues. Report uncertainty, exclusions and disagreements in the answer key.
Human readers and language models should be analysed separately. For model evaluations, record model versions, prompts, sampling settings and independent runs. Multiple outputs from the same model are not equivalent to independent human participants.
Does the wording survive a handoff?
A second phase can pass each version through a short chain of agents. Measure whether an attributed report becomes an unqualified assertion, whether uncertainty is preserved and whether later agents initiate a useful check. The original source and downstream claims need aligned identifiers so that summarisation is not mistaken for fabrication merely because the wording changed.
Research on hallucination snowballing provides related evidence that language models can compound an initial mistake. It does not establish that the four proposed labels prevent that process. This experiment would test a narrower intervention at the point where claims move between speakers.
Failure would be informative
If ordinary copy-editing performs just as well, there is no demonstrated need for a special grammar. If labels make readers overtrust a false verification statement, the intervention may create a new failure. If the benefit disappears after one summary, the design needs support in the handoff format or interface rather than a longer instruction alone.
The paper’s contribution would be a measured answer about communication, accompanied by reusable dialogue materials and an evidence-based answer key. The existing essay contributes the hypothesis. The reader study still has to earn the conclusion.