Brief
A locally run pipeline scores 0.97 and 0.94 on grouping document versions and catching conflicting rules
A new arXiv preprint describes GVD, a system that sorts revised documents into version families and flags rules that duplicate, contradict or refine one another. The reported numbers come from one set of 120 enterprise documents within a single study, so they are a first result rather than a settled one.
Repositories change: guidelines are revised, superseded and re-uploaded, so the same content recurs in different wording. The authors describe GVD, which sorts incoming documents into version families by aligning their rules, then compares each rule against family memory to separate duplicates, contradictions, asymmetric refinements and new knowledge. Counterfactual Span Probing resolves related pairs that inference calls neutral, and relation-specific policies suppress duplicates while escalating consequential changes for review.
The reported test set was 120 enterprise documents processed as 140 ingestions across 59 version families. On that set the framework reached an F1 of 0.97 for version-family construction and 0.94 for rule-level consistency, with Counterfactual Span Probing lifting the second figure from 0.90 to 0.94. Version lineage is kept as an audit trail, and the pipeline runs fully locally, with no large language model.
Our reading
An F1 score is a single number that combines how often a system was right with how often it missed, so these figures describe agreement with whatever standard the authors used, not a guarantee for other document sets. For this desk the value is that the authors name their sample and report numbers instead of anecdotes, which makes the claim checkable. It should matter most to teams who maintain l…
What to do or watch
Watch for the same approach being tested on a second, independent document collection and for a published account of where it fails; until then, treat the single-study numbers as a starting point for evaluation, not a procurement decision.
Source details and supporting facts
Each line is stated by the page named above it.
Stated by arXiv
- On 120 enterprise documents processed as 140 ingestions across 59 version families, GVD reaches an F1 of 0.97 for version-family construction and 0.94 for rule-level consistency.
- Counterfactual Span Probing raises rule consistency from 0.90 to 0.94.
- The pipeline runs fully locally, with no large language model.
- Incoming documents are assigned to version families through bidirectional rule alignment, and their rules are compared against family memory to identify duplicates, contradictions, asymmetric refinements, and new knowledge.
Sources
- arXivText stored 17 September 2026
How this story was checked. Written from the 1 page listed above, stored 17 September 2026; claims checked against that stored text on 17 September 2026.
What that means
- 4 of 4 reported statements were confirmed against the page that carries them; the rest were removed rather than published.
- Figures in the text were required to appear in the stored source text: yes. Identifiers: yes.
- The check reads stored text only: no claim rests on a fresh look that did not happen.
- Where the reporting was silent, the text says so instead of filling the gap.