Source-linked AI summary
GVD: Governed Versioning and Deduplication for Document Repositories
Mohammadreza Sediqin, Shivali Dalmia, Sumukha Thoppanahalli, Abhishek Mukherji
TL;DR
Repositories accumulate revised, overlapping, and sometimes conflicting guidance, while existing work treats versioning, duplicate detection, and contradiction detection as separate pairwise tasks. GVD links incoming documents to version families and governs rule-level updates; on an enterprise corpus, it reports F1 scores of 0.97 for versioning and 0.94 for rule consistency.
Problem
Repositories accumulate revised and conflicting guidance, while existing work handles versioning, duplicate detection, and contradiction detection as isolated pairwise tasks, leaving organizations uncertain which rule version to trust.
Source: §1 · pp. 1–2Method
GVD assigns documents to version families through bidirectional rule alignment, compares rules with family memory, and uses relation-specific update policies, with CSP refining neutral pairs.
Source: §1 · p. 1, §3.2 · p. 4Results
0.97 F1 for versioning and 0.94 for rule consistency on a real enterprise corpus of 120 documents; CSP raises rule-consistency F1 from 0.90 to 0.94.
Source: §5 · p. 7Takeaways & Limitations
Acting on consistency relations rather than stopping at pairwise labels keeps an evolving repository auditable as it grows.
Source: §5 · p. 7Evaluation covers English enterprise guideline documents, and GVD operates on validated rules supplied by an upstream extraction pipeline.
Source: Limitations · p. 8
Abstract
from arXiv · showhide
Document repositories evolve continuously. Guidelines and policies are revised, superseded, and re-uploaded, so the same content recurs in different wording and newer versions refine or contradict earlier ones. These inconsistencies belong to the growing collection rather than to any single document, yet existing work treats versioning, duplicate detection, and contradiction detection as isolated pairwise tasks and stops once a pair is labeled. We present GVD (Governed Versioning and Deduplication), a framework that unifies cross-document version linking with rule-level conflict resolution under an auditable update policy. Incoming documents are assigned to version families through bidirectional rule alignment, and their rules are compared against the family memory to identify duplicates, contradictions, asymmetric refinements, and new knowledge, with Counterfactual Span Probing (CSP) resolving related pairs that inference misclassifies as neutral. Relation-specific policies suppress duplicates and escalate only consequential changes for review, retaining version lineage as an audit trail. The pipeline runs fully locally, with no large language model. 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, with CSP raising rule consistency from 0.90 to 0.94.
1 Introduction
GVD treats repository evolution as a collection-level governance problem: incoming revisions must be linked to history, reconciled with existing rules, and recorded for later inspection. It connects version-family assignment and rule-level consistency decisions to update policies and review.
1 Introduction
- Revisions and re-uploads leave repositories with multiple document versions and overlapping or conflicting guidance, making it harder to know which rule version to trust.In regulated settings, organizations also need to know which version was in force, what replaced it, and who approved the change.Source: §1 · p. 1
- GVD addresses the gap between pairwise classification and governing an evolving repository by linking incoming content to history, reconciling it with existing content, and recording decisions for inspection.The authors describe this as the first framework, to their knowledge, to turn consistency relations into update policy over an evolving collection.Source: §1 · pp. 1–2
- The framework assigns incoming documents to version families through bidirectional rule alignment, then compares their rules with family memory as duplicates, contradictions, asymmetric refinements, or new knowledge.Relation-specific policies suppress duplicates, insert new knowledge, and escalate only consequential changes for review while linking accepted changes to their introducing revision.Source: §1 · p. 1
- Version linking groups documents by bidirectional rule coverage, accommodating revisions that differ in length and wording.Source: §1 · p. 2
- The governed update policy escalates consequential changes for review and retains version lineage as an audit trail, with routing decisions validated against expert review.Source: §1 · p. 2
- Counterfactual Span Probing (CSP) separates cosmetic rewording from substantive edits without training, helping prevent reworded duplicates from entering the collection as new rules.Source: §1 · p. 2
2 Related Work
Prior work addresses document versioning, duplicate detection, contradiction analysis, and human review, but the cited approaches leave repository-level governance of updates unaddressed. GVD extends these lines of work into a repository-maintenance workflow.
2 Related Work
- Lexical fingerprinting can degrade under paraphrasing, while semantic versioning approaches use coarser granularity and do not enforce bidirectional consistency; GVD works at rule level and requires bidirectional agreement.The cited lexical methods include SimHash and Sectional MinHash; semantic approaches include Zhou et al. (2019), Zhu et al. (2017), Sediqin and Argamon (2025b), Spangher et al. (2024b), and Sediqin and Argamon (2025a).Source: §2 · p. 2
- NLI and related pipelines classify duplicate, conflict, or neutral pairs, but do not model whether repository rules should be retained, suppressed, escalated, or added as new.GVD extends semantic classification into a repository-maintenance workflow combining retrieval, bidirectional inference, asymmetric relation detection, and update policies.Source: §2 · p. 2
- Prior work on human review handles document duplication and rule inconsistency in isolation, and cited pipelines stop at classification rather than acting on it; GVD routes consequential cases for validation.The related work names de Sá Baldaia (2020), Surana et al. (2022), Dalmia et al. (2026), and Wu et al. (2022).Source: §2 · p. 2
3 System Framework
GVD governs an evolving repository by linking incoming documents to version families and resolving rule-level relations before updates. Relation-specific policies suppress duplicates and route consequential changes for review while preserving lineage.
3 System Framework
- GVD unifies version-family maintenance and rule-level consistency decisions in an update policy for evolving repositories.It evaluates incoming content before incorporation so the collection stays consistent as it evolves.Source: §3 · p. 2
3.1 Stage 1: Version Maintenance
- Incoming documents join an existing version family only when both directional rule-coverage scores exceed the version threshold.The bidirectional decision guards against absorbing a small set of rules into a larger family when revisions differ in length or wording.Source: §3.1 · p. 3
3.2 Stage 2: Rule Maintenance
- Stage 2 compares incoming rules with the accumulated family memory, using a cosine pre-filter, bidirectional inference, and CSP for residual neutral pairs.Rules below the similarity threshold are inserted directly as new knowledge; only candidate pairs proceed to inference.Source: §3.2 · pp. 3–4
- Counterfactual Span Probing resolves residual neutral rule pairs by distinguishing cosmetic rewordings from substantive edits, refining labels without additional model training.Value or polarity changes are routed for review; a substantive one-sided clause can yield an Asymmetric label, while a two-sided change without those edits remains Neutral.Source: §3.2 · pp. 4–5
3.3 Repository Update and Human Validation
- After classification, duplicates are suppressed, new rules inserted, and contradictions, asymmetric rules, and value or polarity changes routed for human review.Accepted refinements and reviewed edits are linked to parent rules, preserving version lineage and traceability.Source: §3.3 · p. 5
4 Experiments
GVD is evaluated on expert-annotated enterprise documents replayed as repository updates, measuring version grouping, rule consistency, and governance decisions. Results cover benchmark performance, relation handling at scale, review routing, and comparison with prior approaches.
4.1 Dataset and Ground Truth
- The corpus contains 120 enterprise guideline documents replayed as 140 ingestions, forming 59 version families; experts annotated balanced sets of 300 document pairs and 500 rule pairs.The two annotation sets achieved agreement of 0.93 (κ = 0.88) and 0.90 (κ = 0.85), respectively.Source: §4.1 · p. 5
4.2 Experimental Setup
- GVD uses MiniLM embeddings for retrieval and fine-tuned DeBERTa-v3-large for bidirectional inference, with thresholds tuned on a held-out subset disjoint from evaluation sets.Stage 2 metrics are macro-averaged across Duplicate, Contradiction, Asymmetric, and Neutral relations.Source: §4.2 · p. 5
4.3 Versioning and Consistency Results
- F1 reaches 0.971 for version grouping and 0.94 for rule consistency; CSP raises rule-consistency F1 from 0.90 to 0.94.The evaluations use 300 labeled document pairs and 500 labeled rule pairs; CSP's largest gains are on duplicate and neutral classes.Source: §4.3 · p. 5, Table 2 · p. 6, Table 4 · p. 7
4.4 Baselines and Ablations
- For Stage 1, RETSim reaches 0.943 F1 versus GVD's 0.971, while bidirectional coverage beats centroid-only and one-directional matching; Stage 2 CSP lifts F1 to 0.94.Stage 2 baselines include vanilla NLI at 0.72 F1, cosine similarity at 0.56, and bidirectional NLI at 0.90; CSP raises recall from 0.84 to 0.91.Source: §4.4 · p. 6, Table 2 · p. 6, Table 4 · p. 7
4.5 Relation Detection on the Evolving Store
- Across 140 ingestions, GVD made 4,396 rule-level decisions, including 2,719 duplicates, 1,173 new-rule insertions, 194 refinements, 187 contradictions, and 123 neutral cases.Contradictions and asymmetric refinements, including value or polarity changes, yielded 381 cases for human validation.Source: §4.5 · p. 6
4.6 Routing Policy Evaluation
- Reviewers confirmed that routed contradictions, refinements, and value or polarity edits required human judgment; sampled automatic decisions showed no such changes bypassing review.The policy concentrates human effort on consequential updates while automating routine ones.Source: §4.6 · p. 6
4.7 Comparison with Prior Approaches
- Prior systems are summarized on their own datasets rather than compared directly by score; GVD instead unifies rule-level versioning and consistency with repository updates and training-free CSP.The compared approaches either operate at document or sentence level or treat consistency as one-off pairwise classification.Source: §4.7 · pp. 6–7, Table 3 · p. 7
5 Conclusion
GVD governs repository change by linking documents into version families and resolving rule consistency, with consequential changes escalated for human review. On an enterprise corpus, it achieves 0.97 F1 for versioning and 0.94 for rule consistency; the evaluation is scoped to English enterprise guidelines and validated upstream rules.
5 Conclusion
- GVD combines centroid retrieval and bidirectional rule alignment for version families with inference and CSP for rule consistency, escalating consequential changes for review.The pipeline resolves uncertain neutral cases without a large language model and retains lineage as an audit trail.Source: §5 · p. 7
- On a real enterprise corpus of 120 documents, GVD achieves F1 scores of 0.97 for versioning and 0.94 for rule consistency; CSP raises rule-consistency F1 from 0.90 to 0.94.Source: §5 · p. 7
- The evaluation covers English enterprise guideline documents and validated rules supplied by an upstream extraction pipeline; broader domains and languages remain future work.The paper also proposes evaluating larger, more diverse rule sets and integrating GVD into a retrieval-augmented pipeline.Source: §5 · p. 7
Limitations
The evaluation is limited to English enterprise guidelines and governance decisions on validated rules. It does not assess downstream retrieval or generation effects.
Limitations
- The evaluation covers English enterprise guideline documents, with extension to other languages and domains left as future work.The domain choice reflects the availability of expert validation at the scale needed for reliable ground truth.Source: Limitations · p. 8
- GVD operates on validated rules supplied by an upstream extraction pipeline, rather than evaluating rule extraction itself.This keeps versioning and consistency decisions independent of any particular parser.Source: Limitations · p. 8
- The evaluation measures governance decisions; downstream effects on retrieval and generation remain future work.Source: Limitations · p. 8