GVD: Governed Versioning and Deduplication for Document Repositories

# Governed Versioning and Deduplicationfor Document Repositories
Source: [https://arxiv.org/html/2609.17696](https://arxiv.org/html/2609.17696)
## GVD: Governed Versioning and Deduplication for Document Repositories

###### Abstract

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 Guidelines and policies are revised, superseded, and re\-uploaded, so the same content recurs in different wording and newer versions refine or contradict earlier ones\.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\.

## 1Introduction

Document repositories do not stand still\. In enterprise, legal, financial, healthcare, and government settings, guidelines, operational policies, and procedural specifications are revised, superseded, and re\-uploaded as requirements change, so a repository accumulates several versions of the same document alongside overlapping and sometimes conflicting guidance\. Repositories also contain multimodal documents with text, tables, figures, charts, and images[Gao et al\. \(2026\)](https://arxiv.org/html/2609.17696#bib.bib28);[Dalvand et al\. \(2025b\)](https://arxiv.org/html/2609.17696#bib.bib26);[Dalvand et al\. \(2025a\)](https://arxiv.org/html/2609.17696#bib.bib20)\. Duplicates add redundancy, contradictions add ambiguity, and unmanaged revisions make it hard to tell which version of a requirement to trust\. In regulated settings this is not only a quality problem\. An organization needs to know which version of a rule was in force, what replaced it, and who approved the change\.

Existing work treats document similarity, duplicate detection, and contradiction detection as separate pairwise tasks over fixed collections, and stops once a pair has been labeled\. Governing a live repository asks more\. Incoming content must be linked to the history it revises, reconciled with what is already held, and recorded in a way that can be inspected later\. Some of those decisions are routine and some are consequential enough to need a person, and telling the two apart is what keeps review effort affordable as the collection grows\.

We present GVD \(Governed Versioning and Deduplication\), which represents document content as normalized rules and governs the collection in two stages\. Stage 1 assigns each incoming document to a version family through bidirectional rule alignment, so a short revision is not absorbed into a larger family\. Stage 2 compares its rules against the family memory and labels each pair as duplicate, contradiction, asymmetric refinement, or new knowledge\. Relation\-specific policies then act on those labels: duplicates are suppressed, new knowledge is inserted, and only consequential changes are escalated for review, with every accepted change linked to the revision that introduced it\. The pipeline runs locally on encoder\-scale models, with no large language model, so document content never leaves the deployment environment\.

To our knowledge, GVD is the first framework to turn consistency relations into repository update policy over an evolving collection, rather than stopping at classification\. Our contributions are:

1. 1\.A version\-linking method that groups documents into families by bidirectional rule coverage, robust to revisions that differ in length and wording\.
2. 2\.A governed update policy that acts on consistency relations, escalating only consequential changes for review and retaining version lineage as an audit trail, with routing decisions validated against expert review\.
3. 3\.Counterfactual Span Probing \(CSP\), a lightweight training\-free mechanism that separates cosmetic rewording from substantive edits, keeping reworded duplicates from entering the collection as new rules\.

## 2Related Work

Document similarity and versioning\.Lexical fingerprinting methods such as shingling, SimHash\([Williams and Giles, 2013](https://arxiv.org/html/2609.17696#bib.bib1)\), and Sectional MinHash\([Hassanian\-Esfahani and Kargar, 2018](https://arxiv.org/html/2609.17696#bib.bib2)\)degrade under paraphrasing and produce binary duplicate decisions without modeling version families\. Semantic approaches instead apply word\-embedding distances to unsupervised revision detection\([Zhou et al\., 2019](https://arxiv.org/html/2609.17696#bib.bib3);[Zhu et al\., 2017](https://arxiv.org/html/2609.17696#bib.bib4);[Sediqin and Argamon, 2025b](https://arxiv.org/html/2609.17696#bib.bib19)\)and contextual embeddings to sentence\-level alignment across versions\([Spangher et al\., 2024b](https://arxiv.org/html/2609.17696#bib.bib5);[Sediqin and Argamon, 2025a](https://arxiv.org/html/2609.17696#bib.bib27)\), but operate at coarse granularity and do not enforce bidirectional consistency\. GVD operates at the rule level and requires bidirectional agreement\.

Deduplication and contradiction detection\.Because pure similarity cannot distinguish semantic closeness from logical conflict\([Das et al\., 2021](https://arxiv.org/html/2609.17696#bib.bib6);[Shambour et al\., 2022](https://arxiv.org/html/2609.17696#bib.bib7)\), work in requirements engineering turns to natural language inference \(NLI\) to classify pairs as duplicate, conflict, or neutral\([Malik et al\., 2023](https://arxiv.org/html/2609.17696#bib.bib8)\), with hybrid embedding models\([Saleem et al\., 2026](https://arxiv.org/html/2609.17696#bib.bib23)\)and retrieve\-and\-classify pipelines for duplicate bug reports\([Ariai et al\., 2025](https://arxiv.org/html/2609.17696#bib.bib9)\)\. These treat detection as pairwise classification and do not model repository\-level decisions such as whether a rule 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\.

Human\-in\-the\-loop review\.Automatic classification leaves ambiguous cases that require human judgment\([de Sá Baldaia, 2020](https://arxiv.org/html/2609.17696#bib.bib11)\), yet prior work handles document\-level duplication and rule\-level inconsistency in isolation, without governing an evolving collection\([Surana et al\., 2022](https://arxiv.org/html/2609.17696#bib.bib10);[Dalmia et al\., 2026](https://arxiv.org/html/2609.17696#bib.bib21)\), and human\-in\-the\-loop pipelines\([Wu et al\., 2022](https://arxiv.org/html/2609.17696#bib.bib16)\)stop after classification rather than acting on it\. GVD resolves ambiguous cases automatically where possible and routes contradictions, asymmetric refinements, and value or polarity changes for validation, handling routine updates without intervention\.

## 3System Framework

GVD is a governance framework for evolving document repositories\. As documents are revised, re\-uploaded, and extended, the repository accumulates duplicates, conflicting instructions, overlapping guidance, and multiple versions of related processes\. Where existing approaches treat versioning, duplicate detection, and contradiction analysis as independent tasks, GVD unifies them into a single update policy, evaluating incoming content before it is incorporated so the collection stays consistent as it evolves\.

Rather than operating on raw documents, GVD reasons over a shared rule\-level representation, which enables finer\-grained consistency reasoning while reducing sensitivity to formatting and paraphrasing\. Incoming PDF, DOCX, and PPTX documents are first converted intoRuleUnits by an independent multimodal preprocessing pipeline and approved through human\-in\-the\-loop review prior to ingestion; GVD therefore operates only on validated, human\-approved rules and focuses on maintaining consistency within the store rather than generating or validating the rules themselves\. Formally, a document isD=\{𝐫1,𝐫2,…,𝐫N\}D=\\\{\\mathbf\{r\}\_\{1\},\\mathbf\{r\}\_\{2\},\\ldots,\\mathbf\{r\}\_\{N\}\\\}where each rule embedding𝐫i∈ℝ384\\mathbf\{r\}\_\{i\}\\in\\mathbb\{R\}^\{384\}is generated using the all\-MiniLM\-L6\-v2 sentence encoder, which provides a lightweight 384\-dimensional representation while maintaining strong semantic retrieval performance\. The same encoder is used throughout both stages to ensure a consistent embedding space for document retrieval, version matching, and rule\-level consistency analysis, as illustrated in Figure[1](https://arxiv.org/html/2609.17696#S3.F1)\.

![Refer to caption](https://arxiv.org/html/2609.17696v1/arc4.png)Figure 1:Overview of GVD\. Incoming documents are converted into validated RuleUnits and processed in two stages: \(1\) version maintenance assigns documents to families, and \(2\) rule maintenance classifies rules into consistency relations via bidirectional NLI and CSP\. Relation\-specific policies then update the store, suppressing duplicates automatically and routing consequential changes for human review\.### 3\.1Stage 1: Version Maintenance

The objective of this stage is to determine whether an incoming document belongs to an existing version family or represents new knowledge that should initialize a new family\. This is non\-trivial because successive versions of the same guideline document often differ substantially in content and rule count: a new revision may add, remove, or rephrase rules, so two documents in the same family need not contain the same number of rules or identical text\. Stage 1 therefore proceeds in two steps, a fast centroid\-based retrieval followed by a stricter bidirectional alignment\.

Stage 1A: Centroid retrieval\.Each document is summarized by the centroid of its rule embeddings,

𝐜D=1\|D\|​∑i=1\|D\|𝐫i,\\mathbf\{c\}\_\{D\}=\\frac\{1\}\{\|D\|\}\\sum\_\{i=1\}^\{\|D\|\}\\mathbf\{r\}\_\{i\},\(1\)where\|D\|\|D\|is the number of rules in the document\. For an incoming documentDqD\_\{q\}and a stored documentDjD\_\{j\}, centroid similarity is computed by cosine similarity,

sim⁡\(Dq,Dj\)=𝐜Dq⋅𝐜Dj∥𝐜Dq∥​∥𝐜Dj∥\.\\mathrm\{sim\}\(D\_\{q\},D\_\{j\}\)=\\frac\{\\mathbf\{c\}\_\{D\_\{q\}\}\\cdot\\mathbf\{c\}\_\{D\_\{j\}\}\}\{\\lVert\\mathbf\{c\}\_\{D\_\{q\}\}\\rVert\\,\\lVert\\mathbf\{c\}\_\{D\_\{j\}\}\\rVert\}\.\(2\)The top\-kkdocuments exceeding a retrieval threshold are retained as candidates, shrinking the search space before the more expensive rule\-level comparison\.

Stage 1B: Bidirectional alignment\.Each candidate then undergoes rule\-level alignment, where the similarity between two rules is

s⁡\(𝐫i,𝐫j\)=𝐫i⋅𝐫j∥𝐫i∥​∥𝐫j∥\.s\(\\mathbf\{r\}\_\{i\},\\mathbf\{r\}\_\{j\}\)=\\frac\{\\mathbf\{r\}\_\{i\}\\cdot\\mathbf\{r\}\_\{j\}\}\{\\lVert\\mathbf\{r\}\_\{i\}\\rVert\\,\\lVert\\mathbf\{r\}\_\{j\}\\rVert\}\.\(3\)Forward coverage measures how many rules in the incoming document are represented in the candidate,

Sf=1\|Dq\|​∑𝐫i∈Dq𝕀⁡\(max𝐫j∈Dc⁡s⁡\(𝐫i,𝐫j\)≥τ\),S\_\{f\}=\\frac\{1\}\{\|D\_\{q\}\|\}\\sum\_\{\\mathbf\{r\}\_\{i\}\\in D\_\{q\}\}\\mathbb\{I\}\\\!\\left\(\\max\_\{\\mathbf\{r\}\_\{j\}\\in D\_\{c\}\}s\(\\mathbf\{r\}\_\{i\},\\mathbf\{r\}\_\{j\}\)\\geq\\tau\\right\),\(4\)while reverse coverage measures the converse,

Sr=1\|Dc\|​∑𝐫j∈Dc𝕀⁡\(max𝐫i∈Dq⁡s⁡\(𝐫i,𝐫j\)≥τ\)\.S\_\{r\}=\\frac\{1\}\{\|D\_\{c\}\|\}\\sum\_\{\\mathbf\{r\}\_\{j\}\\in D\_\{c\}\}\\mathbb\{I\}\\\!\\left\(\\max\_\{\\mathbf\{r\}\_\{i\}\\in D\_\{q\}\}s\(\\mathbf\{r\}\_\{i\},\\mathbf\{r\}\_\{j\}\)\\geq\\tau\\right\)\.\(5\)The incoming document joins an existing family through a dual\-coverage decision: it joins only when both directional coverage scores exceed the version threshold,

\(Dq,Dc\)∈F⇔Sf≥τv∧Sr≥τv\.\(D\_\{q\},D\_\{c\}\)\\in F\\iff S\_\{f\}\\geq\\tau\_\{v\}\\;\\land\\;S\_\{r\}\\geq\\tau\_\{v\}\.\(6\)Requiring agreement in both directions, rather than one\-directional similarity, prevents a small set of rules from being wrongly absorbed into a larger family and gives a more reliable picture of document evolution even when versions differ in length or wording\.

### 3\.2Stage 2: Rule Maintenance

After version assignment, the incoming document enters its version family and its rules are compared against the family memory, which consists of all validated rules previously associated with that lineage, including rules from earlier revisions, historical versions, and accepted repository updates\. Consistency decisions are therefore made not only against the current document but against the accumulated knowledge of the entire version family\. Stage 2 proceeds in three steps: a cosine similarity pre\-filter that retrieves candidate rules, bidirectional inference that assigns a consistency label, and Counterfactual Span Probing that resolves the residual neutral pairs\.

Stage 2A: Similarity pre\-filter\.For each incoming rule, cosine similarity identifies the most relevant candidate rule in the family memory\. A rule whose best cosine similarity falls below a thresholdτs\\tau\_\{s\}has no close match in the store and is inserted directly as new knowledge; only pairs at or aboveτs\\tau\_\{s\}proceed to inference\. We setτs=0\.85\\tau\_\{s\}=0\.85\.

Stage 2B: Bidirectional inference\.For each candidate pair, inference is performed with DeBERTa\-v3\-large fine\-tuned on SNLI and MultiNLI, applied in both directions\. LetPe→P\_\{e\}^\{\\rightarrow\}andPe←P\_\{e\}^\{\\leftarrow\}denote the entailment probabilities of the existing\-to\-new and new\-to\-existing directions, andPcP\_\{c\}the contradiction probability\. The pair is labeled as

Label=\{Contradiction,max⁡\(Pc→,Pc←\)≥τc,Duplicate,Pe→≥τd∧Pe←≥τd,Asymmetric,\(Pe→≥τd\)⊕\(Pe←≥τd\),Neutral,Pe→<τd∧Pe←<τd,\\mathrm\{Label\}=\{\\begin\{cases\}\\text\{Contradiction\},&\\max\(P\_\{c\}^\{\\rightarrow\},P\_\{c\}^\{\\leftarrow\}\)\\geq\\tau\_\{c\},\\\\\[2\.0pt\] \\text\{Duplicate\},&P\_\{e\}^\{\\rightarrow\}\\geq\\tau\_\{d\}\\land P\_\{e\}^\{\\leftarrow\}\\geq\\tau\_\{d\},\\\\\[2\.0pt\] \\text\{Asymmetric\},&\(P\_\{e\}^\{\\rightarrow\}\\geq\\tau\_\{d\}\)\\oplus\(P\_\{e\}^\{\\leftarrow\}\\geq\\tau\_\{d\}\),\\\\\[2\.0pt\] \\text\{Neutral\},&P\_\{e\}^\{\\rightarrow\}<\\tau\_\{d\}\\land P\_\{e\}^\{\\leftarrow\}<\\tau\_\{d\},\\end\{cases\}\}\(7\)where→\\rightarrowand←\\leftarrowdenote the directionsrexist→rnewr\_\{\\text\{exist\}\}\\\!\\rightarrow\\\!r\_\{\\text\{new\}\}andrnew→rexistr\_\{\\text\{new\}\}\\\!\\rightarrow\\\!r\_\{\\text\{exist\}\}, and⊕\\oplusdenotes exclusive or\. A pair is a Duplicate when entailment holds in both directions, Asymmetric when it holds in exactly one direction, capturing the refinements and partial extensions that frequently occur as documents evolve, and Neutral when neither direction entails and there is no contradiction\.

Stage 2C: Counterfactual Span Probing\.Bidirectional inference reliably identifies clear duplicates and contradictions, but many semantically related rules fall into the Neutral class, where hidden duplicates and asymmetric refinements remain unresolved\. CSP is a training\-free refinement applied only to neutral pairs, designed to recover latent duplicate and asymmetric relations that single\-pass inference collapses into the neutral class\. Rather than re\-scoring entire rules, CSP isolates the spans responsible for disagreement and probes, counterfactually, whether those differences are substantive or merely cosmetic\. This allows GVD to distinguish cosmetic reformulations from substantive modifications without additional model training\. The identified spans are categorized as cosmetic rewordings, value or numerical changes, polarity reversals, or one\-sided clauses, which are used to refine the final consistency label\. LetVV,PP,C1C\_\{1\}, andC2C\_\{2\}denote, respectively, the presence of any value or numerical change, any polarity reversal, a substantive one\-sided clause on exactly one side, and substantive clauses on both sides\. The pair is relabeled as

CSP=\{Review,V∨P,Neutral,¬\(V∨P\)∧C2,Asymmetric,¬\(V∨P∨C2\)∧C1,Duplicate,otherwise\.\\mathrm\{CSP\}=\\begin\{cases\}\\text\{Review\},&V\\lor P,\\\\\[2\.0pt\] \\text\{Neutral\},&\\lnot\(V\\lor P\)\\land C\_\{2\},\\\\\[2\.0pt\] \\text\{Asymmetric\},&\\lnot\(V\\lor P\\lor C\_\{2\}\)\\land C\_\{1\},\\\\\[2\.0pt\] \\text\{Duplicate\},&\\text\{otherwise\}\.\\end\{cases\}\(8\)A pair is labeled Asymmetric only when a substantive one\-sided clause is its sole non\-cosmetic difference; a value or polarity change is instead routed to human review as a substantive edit to an existing rule, while a two\-sided change with no value or polarity edit remains Neutral\. A one\-sided clause is substantive when, after discarding short function\-like tokens \(fewer than three characters, such as “if”, “to”, or “a”\), at least three content tokens remain\. Both thresholds were set on the validation subset: lower values admit trivial additions as spurious refinements, while higher values discard genuine clause\-level changes\. The inference model is re\-applied only on replacement spans, as a targeted polarity and contradiction check, rather than as the labeling rule; the Duplicate and Asymmetric decisions themselves rest on span structure and token counts and require no further inference call\. CSP operates only on residual neutral pairs, routing value and polarity edits to review and otherwise emitting Duplicate, Asymmetric, or Neutral\. Contradictions are detected upstream by Eq\.[7](https://arxiv.org/html/2609.17696#S3.E7), which classifies them reliably \(F1 0\.97\)\.

Table 1:Two CSP reclassification examples\.Table[1](https://arxiv.org/html/2609.17696#S3.T1)shows two such cases: a cosmetic rewording relabeled Duplicate, and a substantive added clause relabeled Asymmetric\.

### 3\.3Repository Update and Human Validation

Following consistency classification, repository updates follow relation\-specific policies\. Duplicate rules are suppressed, and new rules are inserted directly\. Contradictory rules, asymmetric rules, and rules carrying a value or polarity change are quarantined and routed for human review before repository modification, while asymmetric and reviewed edits are stored as refinements linked to their parent rule, preserving version lineage while preventing conflicting values from silently entering the store\. By combining rule\-level representation, version\-family maintenance, bidirectional consistency reasoning, CSP, and selective human validation within a single framework, GVD governs how a document repository changes over time while keeping every accepted update traceable to the revision that introduced it\.

## 4Experiments

### 4\.1Dataset and Ground Truth

Because evolving enterprise repositories with expert validation are hard to obtain at scale, we prioritize a real, expertly annotated corpus over a larger synthetic one\. Our evaluation corpus consists of 120 enterprise guideline documents provided by industrial clients under confidentiality agreements, in PDF, DOCX, and PPTX format, processed into approximately 3,900 validated RuleUnits\. The corpus was replayed as 140 repository ingestions to simulate evolution: 59 first uploads initializing version families, 61 revised versions of existing documents, and 20 exact reuploads, forming 59 version families and yielding 4,396 rule\-level consistency decisions\. Since enterprise guideline repositories provide no ground\-truth labels, we construct balanced evaluation sets through expert human annotation\. For Stage 1, we sample 300 document pairs across the 59 version families, balanced into 150 same\-family and 150 different\-family pairs, ensuring balanced evaluation of grouping and separation\. For Stage 2, we annotate 500 rule pairs across four consistency relations \(130 duplicate, 122 contradiction, 128 asymmetric, 120 neutral\)\. Each set was independently labeled by expert annotators familiar with the guidelines, achieving 0\.93 agreement with Cohen’sκ=0\.88\\kappa=0\.88\(Stage 1\) and 0\.90 agreement withκ=0\.85\\kappa=0\.85\(Stage 2\)\.

### 4\.2Experimental Setup

GVD uses the all\-MiniLM\-L6\-v2 sentence encoder to produce rule embeddings for both document\-level and rule\-level retrieval, and DeBERTa\-v3\-large fine\-tuned for Natural Language Inference for bidirectional semantic inference\. All thresholds were tuned on a small held\-out validation subset, disjoint from the Stage 1 and Stage 2 evaluation sets to avoid leakage, and then fixed for all reported experiments\. For Stage 1A candidate retrieval, we evaluatedk∈\{3,5,8,10\}k\\in\\\{3,5,8,10\\\}on this validation subset and fixedk=5k=5;k=3k=3reduced family coverage, whilek=8k=8andk=10k=10added only redundant candidates that bidirectional alignment subsequently removed\. For Stage 1, we set the centroid thresholdτr=0\.70\\tau\_\{r\}=0\.70, per\-rule matchτ=0\.75\\tau=0\.75, and version thresholdτv=0\.75\\tau\_\{v\}=0\.75; for Stage 2,τs=0\.85\\tau\_\{s\}=0\.85, entailmentτd=0\.70\\tau\_\{d\}=0\.70, and contradictionτc=0\.85\\tau\_\{c\}=0\.85\. All experiments run on a single NVIDIA GPU\. Stage 1 is a binary decision, reported as precision, recall, and F1\. For Stage 2, the precision, recall, and F1 values reported in Table[5](https://arxiv.org/html/2609.17696#S4.T5)are macro\-averaged across the four consistency relations \(Duplicate, Contradiction, Asymmetric, and Neutral\)\.

### 4\.3Versioning and Consistency Results

On the 300 labeled document pairs, GVD reaches an F1 of 0\.971 for version grouping \(Table[4](https://arxiv.org/html/2609.17696#S4.T4)\), correctly placing same\-family documents together and keeping unrelated ones apart\. On the 500 labeled rule pairs, it reaches precision 0\.97, recall 0\.91, and F1 0\.94 \(Table[5](https://arxiv.org/html/2609.17696#S4.T5)\)\. Bidirectional inference handles clear duplicates and contradictions well but leaves many related rules in the neutral class, where hidden duplicates and refinements would otherwise enter the store as new\. CSP revisits these neutral pairs and raises rule\-consistency F1 from 0\.90 to 0\.94, with the largest gains on the duplicate and neutral classes \(Table[2](https://arxiv.org/html/2609.17696#S4.T2)\)\.

Table 2:Rule\-relation classification: effect of CSP, per class and overall\.
### 4\.4Baselines and Ablations

Because no prior system directly targets this task, we evaluate standard baselines and ablations on our labeled datasets\. For Stage 1, we compare lexical fingerprinting methods \(MinHash and Shingling with Jaccard\([Broder, 1997](https://arxiv.org/html/2609.17696#bib.bib18)\), and SimHash\([Charikar, 2002](https://arxiv.org/html/2609.17696#bib.bib22)\)\) and embedding\-based similarity \(SBERT cosine and clustering\([Reimers and Gurevych, 2019](https://arxiv.org/html/2609.17696#bib.bib12)\)\)\. We additionally compare against RETSim\([Zhang et al\., 2024](https://arxiv.org/html/2609.17696#bib.bib25)\), a lightweight model purpose\-built for near\-duplicate and versioning detection, which embeds text with a character\-level vectorizer trained via metric learning and compares embeddings by cosine similarity\. For all score\-based baselines, the same\-family decision threshold was selected on the same held\-out validation subset used to tune GVD’s thresholds, so that each method is evaluated at its own best operating point\. The lexical and general\-embedding baselines reach only 0\.62 to 0\.70 F1 \(SBERT cosine strongest at 0\.704\)\. RETSim is far stronger, reaching 0\.943, reflecting its specialization for near\-duplicate detection\. It nonetheless still trails GVD \(0\.971\): because RETSim scores document similarity directly, it does not model the bidirectional, family\-level coverage that reliably separates same\-family from different\-family documents, particularly when versions differ in length or wording\. The ablation then isolates the matching rule: centroid\-only similarity is too coarse \(0\.837\), one\-directional coverage still over\-merges short documents \(0\.944\), and bidirectional coverage gives the best result \(0\.971\) \(Table[4](https://arxiv.org/html/2609.17696#S4.T4)\)\.

Table 3:Prior systems on their own datasets, not directly comparable to GVD\.For Stage 2 \(Table[5](https://arxiv.org/html/2609.17696#S4.T5)\), a vanilla NLI baseline that maps a single forward pass directly to a label reaches 0\.72\. The component ablation isolates each part: cosine similarity alone reaches only 0\.56, since it cannot separate logical relations from surface closeness; adding bidirectional NLI raises F1 to 0\.90; and CSP lifts recall from 0\.84 to 0\.91 \(F1 0\.94\) by recovering hidden duplicates and asymmetric refinements left in the neutral class\. Lexical methods such as MinHash and SimHash apply only to Stage 1, as they yield similarity scores and cannot express contradiction or asymmetric relations; for Stage 2 we therefore compare against cosine\-only and vanilla NLI baselines\.

Table 4:Stage 1 version grouping \(300 pairs\)\.Table 5:Stage 2 rule consistency \(500 pairs\)\.
### 4\.5Relation Detection on the Evolving Store

Across all 140 document ingestions, each incoming rule is either inserted directly as new knowledge when no sufficiently similar rule exists in the family memory, or compared against its nearest retrieved candidate and assigned a consistency relation\. In total, GVD processed 4,396 rule\-level decisions, yielding 2,719 duplicate relations, 1,173 new\-rule insertions, 194 asymmetric refinements, 187 contradictions, and 123 neutral cases after CSP refinement\. Duplicate relations dominate, reflecting the substantial overlap that naturally occurs across document revisions and reuploads\. Contradictions and asymmetric refinements, including those carrying a value or polarity change, are routed for human validation, giving 381 cases for review\. The remaining neutral cases are genuinely new and inserted directly\.

### 4\.6Routing Policy Evaluation

To evaluate the governance policy rather than relation classification alone, we manually reviewed the cases routed for human validation, comprising contradictions, asymmetric refinements, and value or polarity edits\. Reviewers confirmed that these cases required human judgment before repository modification, and a sample of automatically handled decisions showed no instance in which a contradiction, refinement, or value or polarity change bypassed review\. The governance layer therefore concentrates human effort on consequential updates while automating routine ones\.

### 4\.7Comparison with Prior Approaches

Table[3](https://arxiv.org/html/2609.17696#S4.T3)lists prior systems for versioning, deduplication, and contradiction detection\. Because they are evaluated on different datasets and tasks, we do not compare their scores directly against GVD; we instead summarize each method and the result it reports on its own benchmark\. For document versioning and near\-duplicate detection, NewsEdits\([Spangher et al\., 2024a](https://arxiv.org/html/2609.17696#bib.bib13)\)aligns sentences across document versions with contextual embeddings to detect revisions, improving cross\-version sentence linking over its earlier version\. arXivEdits\([Jiang et al\., 2022](https://arxiv.org/html/2609.17696#bib.bib24)\)trains a neural CRF to align sentences across multiple revisions of scientific papers, reaching high alignment F1 on its own revision corpus\. For rule\-level consistency, FSARC\([Guo et al\., 2021](https://arxiv.org/html/2609.17696#bib.bib14)\)parses requirements into a semantic eight\-tuple and applies heuristic rules to flag conflicts, achieving high recall on natural\-language requirements without supervised training\.[Malik et al\. \(2023\)](https://arxiv.org/html/2609.17696#bib.bib8)use sequential transfer learning over transformer encoders to classify requirement pairs as duplicate, conflict, or neutral, improving accuracy in data\-rich settings\. PassionNet\([Saleem et al\., 2025](https://arxiv.org/html/2609.17696#bib.bib17)\)combines an LLM with similarity features to identify duplicate and conflicting requirements, while LegalWiz\([Mantravadi et al\., 2025](https://arxiv.org/html/2609.17696#bib.bib15)\)pairs NLI with an LLM judge to detect contradictions in legal text\.

These systems share two limitations that GVD addresses: versioning and near\-duplicate methods operate only at the document or sentence level and cannot express contradiction or asymmetric refinement, while consistency methods treat detection as one\-off pairwise classification without modeling repository\-level updates over an evolving store\. GVD works at the rule level, unifies versioning and consistency within a single pipeline, and resolves ambiguous neutral pairs through training\-free CSP, which recovers latent relations missed by single\-pass inference without relying on a large language model\.

## 5Conclusion

We presented GVD, a fully local framework for governing how a document repository changes as it accumulates revisions\. GVD groups documents into version families through centroid retrieval and bidirectional rule alignment, and manages rule\-level consistency through inference and Counterfactual Span Probing, escalating only consequential changes for human review and resolving uncertain neutral cases without any large language model\. On a real enterprise corpus of 120 documents, it reaches an F1 of 0\.97 for versioning and 0\.94 for rule consistency, with CSP raising F1 from 0\.90 to 0\.94\. Our results show that governing an evolving collection is a distinct problem from pairwise duplicate or contradiction detection, and that acting on consistency relations rather than stopping at them keeps the repository auditable as it grows\. In future work, we plan to extend GVD across enterprise domains and languages, evaluate it on larger and more diverse rule sets, and integrate it into a retrieval\-augmented pipeline to measure its effect on downstream answer quality\.

## Limitations

The present evaluation covers English enterprise guideline documents, chosen because expert validation was available at the scale required for reliable ground truth; extension to further languages and document domains is a natural next step\. GVD is scoped to the governance layer and operates on validated rules supplied by an upstream extraction pipeline, which keeps versioning and consistency decisions independent of any particular parser\. It is evaluated on the governance decisions themselves, through version\-family construction, relation classification, and routing validation, so its effect on downstream retrieval and generation is a natural direction for future work\.

## References

- Ariaiet al\.\(2025\)F\. Ariai, J\. Mackenzie, and G\. DemartiniNatural language processing for the legal domain: a survey of tasks, datasets, models, and challenges\.ACM Computing Surveys58\(6\),pp\. 1–37\.Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p2.1)\.
- Broder \(1997\)A\. Z\. BroderOn the resemblance and containment of documents\.InProceedings\. Compression and Complexity of SEQUENCES 1997 \(Cat\. No\. 97TB100171\),pp\. 21–29\.Cited by:[§4\.4](https://arxiv.org/html/2609.17696#S4.SS4.p1.1)\.
- Charikar \(2002\)M\. S\. CharikarSimilarity estimation techniques from rounding algorithms\.InProceedings of the Thirty\-Fourth Annual ACM Symposium on Theory of Computing,pp\. 380–388\.Cited by:[§4\.4](https://arxiv.org/html/2609.17696#S4.SS4.p1.1)\.
- Dalmiaet al\.\(2026\)S\. Dalmia, S\. Thoppanahalli, M\. Sediqin, and A\. MukherjiGUIDE: governed unified intelligence for document\-to\-artifact generation in enterprise settings\.arXiv preprint arXiv:2608\.12133\.Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p3.1)\.
- Dalvandet al\.\(2025a\)F\. Dalvand, A\. Dutkiewicz, N\. M\. Wright, B\. R\. Mather, and D\. MüllerRegional carbonate compensation depth variability in the pacific ocean since the oligocene\.Frontiers in Earth Science13,pp\. 1605906\.Cited by:[§1](https://arxiv.org/html/2609.17696#S1.p1.1)\.
- Dalvandet al\.\(2025b\)F\. Dalvand, A\. Dutkiewicz, N\. M\. Wright, and R\. D\. MüllerIndian ocean carbonate compensation depth since the late oligocene\.Geo\-Marine Letters45\(4\),pp\. 38\.Cited by:[§1](https://arxiv.org/html/2609.17696#S1.p1.1)\.
- Daset al\.\(2021\)S\. Das, N\. Deb, A\. Cortesi, and N\. ChakiSentence embedding models for similarity detection of software requirements\.SN Computer Science2\(2\),pp\. 69\.Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p2.1)\.
- de Sá Baldaia \(2020\)B\. S\. de Sá BaldaiaLie\-o\-matic: using natural language processing to detect contradictory statements\.Master’s Thesis,Universidade do Porto \(Portugal\)\.Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p3.1)\.
- Gaoet al\.\(2026\)S\. Gao, S\. Zhao, X\. Jiang, L\. Duan, Y\. X\. Chng, Q\. Chen, W\. Luo, K\. Zhang, J\. Bian, and M\. GongScaling beyond context: a survey of multimodal retrieval\-augmented generation for document understanding\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 4458–4489\.Cited by:[§1](https://arxiv.org/html/2609.17696#S1.p1.1)\.
- Guoet al\.\(2021\)W\. Guo, L\. Zhang, and X\. LianAutomatically detecting the conflicts between software requirements based on finer semantic analysis\.arXiv preprint arXiv:2103\.02255\.Cited by:[§4\.7](https://arxiv.org/html/2609.17696#S4.SS7.p1.1)\.
- Hassanian\-Esfahani and Kargar \(2018\)R\. Hassanian\-Esfahani and M\. KargarSectional minhash for near\-duplicate detection\.Expert Systems with Applications99,pp\. 203–212\.Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p1.1)\.
- Jianget al\.\(2022\)C\. Jiang, W\. Xu, and S\. StevensarXivEdits: understanding the human revision process in scientific writing\.InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),Abu Dhabi, United Arab Emirates,pp\. 9420–9435\.External Links:[Document](https://dx.doi.org/10.18653/v1/2022.emnlp-main.641),[Link](https://aclanthology.org/2022.emnlp-main.641/)Cited by:[§4\.7](https://arxiv.org/html/2609.17696#S4.SS7.p1.1)\.
- Maliket al\.\(2023\)G\. Malik, S\. Yildirim, M\. Cevik, A\. Bener, and D\. ParikhTransfer learning for conflict and duplicate detection in software requirement pairs\.arXiv preprint arXiv:2301\.03709\.Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p2.1),[§4\.7](https://arxiv.org/html/2609.17696#S4.SS7.p1.1)\.
- Mantravadiet al\.\(2025\)A\. Mantravadi, S\. Dalmia, O\. Pospelova, A\. Mukherji, N\. Dave, and A\. MittalLegalWiz: a multi\-agent generation framework for contradiction detection in legal documents\.arXiv preprint arXiv:2510\.03418\.Cited by:[§4\.7](https://arxiv.org/html/2609.17696#S4.SS7.p1.1)\.
- Reimers and Gurevych \(2019\)N\. Reimers and I\. GurevychSentence\-bert: sentence embeddings using siamese bert\-networks\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing \(EMNLP\-IJCNLP\),pp\. 3982–3992\.Cited by:[§4\.4](https://arxiv.org/html/2609.17696#S4.SS4.p1.1)\.
- Saleemet al\.\(2025\)S\. Saleem, M\. N\. Asim, and A\. DengelPassionnet: an innovative framework for duplicate and conflicting requirements identification\.Expert Systems with Applications293,pp\. 128684\.Cited by:[§4\.7](https://arxiv.org/html/2609.17696#S4.SS7.p1.1)\.
- Saleemet al\.\(2026\)S\. Saleem, M\. N\. Asim, and A\. DengelReqNet: an llm\-driven computational framework for automated requirements extraction from unstructured documents\.Complex & Intelligent Systems12\(1\),pp\. 38\.Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p2.1)\.
- Sediqin and Argamon \(2025a\)M\. Sediqin and S\. E\. ArgamonLACES: lexical and contextual awareness for edu segmentation\.InInternational Conference on Natural Language Processing and Information Retrieval,pp\. 47–60\.Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p1.1)\.
- Sediqin and Argamon \(2025b\)M\. Sediqin and S\. E\. ArgamonRST\-ndas:rst nuclei for discourse\-aware abstractive summarization\.In2025 IEEE 12th International Conference on Data Science and Advanced Analytics \(DSAA\),Vol\.,pp\. 1–5\.External Links:[Document](https://dx.doi.org/10.1109/DSAA65442.2025.11248014)Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p1.1)\.
- Shambouret al\.\(2022\)Q\. Y\. Shambour, A\. H\. Hussein, Q\. M\. Kharma, and M\. M\. AbualhajEffective hybrid content\-based collaborative filtering approach for requirements engineering\.Computer Systems Science & Engineering40\(1\)\.Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p2.1)\.
- Spangheret al\.\(2024a\)A\. Spangher, K\. Huang, H\. Cho, and J\. MayNewsEdits 2\.0: learning the intentions behind updating news\.arXiv preprint arXiv:2411\.18811\.Cited by:[§4\.7](https://arxiv.org/html/2609.17696#S4.SS7.p1.1)\.
- Spangheret al\.\(2024b\)A\. Spangher, N\. Peng, S\. Gehrmann, and M\. DredzeDo llms plan like human writers? comparing journalist coverage of press releases with llms\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,pp\. 21814–21828\.Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p1.1)\.
- Suranaet al\.\(2022\)S\. Surana, S\. Dembla, and P\. BihaniIdentifying contradictions in the legal proceedings using natural language models\.SN Computer Science3\(3\),pp\. 187\.Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p3.1)\.
- Williams and Giles \(2013\)K\. Williams and C\. L\. GilesNear duplicate detection in an academic digital library\.InProceedings of the 2013 ACM symposium on Document engineering,pp\. 91–94\.Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p1.1)\.
- Wuet al\.\(2022\)X\. Wu, L\. Xiao, Y\. Sun, J\. Zhang, T\. Ma, and L\. HeA survey of human\-in\-the\-loop for machine learning\.Future Generation Computer Systems135,pp\. 364–381\.Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p3.1)\.
- Zhanget al\.\(2024\)M\. Zhang, O\. Vallis, A\. Bumin, T\. Vakharia, and E\. BurszteinRETSim: resilient and efficient text similarity\.InInternational Conference on Learning Representations \(ICLR\),External Links:[Link](https://arxiv.org/abs/2311.17264)Cited by:[§4\.4](https://arxiv.org/html/2609.17696#S4.SS4.p1.1)\.
- Zhouet al\.\(2019\)S\. Zhou, X\. Xu, Y\. Liu, R\. Chang, and Y\. XiaoText similarity measurement of semantic cognition based on word vector distance decentralization with clustering analysis\.IEEE Access7,pp\. 107247–107258\.Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p1.1)\.
- Zhuet al\.\(2017\)X\. Zhu, D\. Klabjan, and P\. BlessSemantic document distance measures and unsupervised document revision detection\.InProceedings of the Eighth International Joint Conference on Natural Language Processing \(Volume 1: Long Papers\),pp\. 947–956\.Cited by:[§2](https://arxiv.org/html/2609.17696#S2.p1.1)\.