Checking whether a source still supports a claim
A citation is useful only while its supporting context remains inspectable.
A note may link to a document that later changes. The link can still work while the relevant date, amount, qualification or business scope has changed. Doppel has a bounded review path for comparing the earlier and current source.
“Batch jobs run in the EU region.”
“Batch jobs now run in the US region.”
Begin with the original.
The comparison starts from an accepted citation and its earlier source version. Doppel checks source identity and scope before asking a model about meaning.
The earlier excerpt must have a unique context in the original. The comparison includes its section heading and full passage. Removed wording or a missing baseline produces an explicit review result rather than a reconstructed source.
This preserves the question being asked: what changed in the material that supported this citation?
Find the relevant revision.
The code compares earlier and current source units, keeping section headings with their paragraphs. It narrows changed material using the cited section and shared content terms.
If the cited passage and section are intact and edits are confined elsewhere, the result can be unchanged without a model call. Removing qualifying wording can mark the source as affected. A passage from a different section needs scope review even if it uses familiar names.
These rules reduce the chance that unrelated or incomplete context is treated as supporting evidence.
Compare meaning in both directions.
For eligible changed passages, an on-device NLI model evaluates whether the current wording supports the earlier meaning and whether the earlier wording supports the current meaning.
A one-way support score can miss an added condition or a change in specificity. The implementation requires agreement in both directions, low contradiction scores, and preservation of protected details such as numbers, names and qualifiers.
The comparison has a fixed input and pair budget. Missing model output, overlength material or unavailable source context leaves the check incomplete.
Keep the result reviewable.
| Result | Meaning in this review path |
|---|---|
| Unchanged | The relevant source passage and section remain intact under the implemented comparison rules. |
| Equivalent | Eligible changed passages pass the bidirectional meaning check and preserve protected details. |
| Affected | The supporting wording may have changed or lost a qualification. Inspect the current passage. |
| Incomplete | The complete comparison could not be made within the available source, model or scope limits. |
The result concerns support in a source. It does not establish that the source itself is factually correct, or authorize rewriting the user’s note. Current passages and the reason for the result remain part of the review.
Test the distinctions that matter.
The repository’s focused fixtures cover changes to supporting meaning, preserved qualifications, scope changes and unavailable comparisons. Those tests protect the implementation contract; they do not establish accuracy across every kind of prose.
A broader quality evaluation should hold out complete documents and revision families. Report false “equivalent” results alongside how many cases the system can resolve. A system that quietly skips difficult cases needs its incomplete coverage reported too.
The useful product behavior is a traceable review: the earlier support, the current wording, the reason for concern and a clear indication of what the check could establish.
Doppel implements this source-comparison path with a ModernBERT NLI model and explicit source, scope and detail checks. The examples here are illustrative. This note reports no new independent accuracy benchmark.
See the product context.
Doppel applies source-aware checks alongside local inference, hybrid retrieval and meeting review.
Read the Doppel case study