OEP-Bench

OEP-Bench (Orex Enterprise Policy Benchmark) is a synthetic benchmark for evaluating retrieval-augmented generation (RAG) and enterprise knowledge systems under policy versioning, access-control, citation, temporal-reasoning, and security constraints.

OEP-Bench is built around NovaCore Technologies, a fictional enterprise used solely as the organizational setting for the benchmark. NovaCore Technologies is not the publisher or owner of the benchmark.

The benchmark contains 884 evaluation cases across 40 policy families and 8 organizational domains. Its retrievable document corpus contains 121 document versions:

• 119 canonical main-policy versions
• 2 jurisdiction-specific addenda associated with the E1 Data Privacy policy family

Of the 40 main policy families, 39 contain versions 1.0, 2.0, and 3.0. The E5 Whistleblower policy intentionally contains only versions 1.0 and 2.0.

OEP-Bench is designed to study enterprise retrieval and RAG behavior where relevance alone is insufficient. Depending on the case, a system may also need to identify the policy version applicable at a particular time, respect document permissions, distinguish authoritative evidence from confusable material, produce version-correct citations, or decline to provide an answer when appropriate.

Benchmark at a Glance

Component	Count
Organizational domains	8
Policy families	40
Canonical main-policy versions	119
E1 addenda	2
Retrievable document versions	121
Evaluation cases	884
Ground-truth records	884
Permission records	121
Per-family metadata PDFs	40
Per-family permission PDFs	40
Per-family question PDFs	40
Per-family ground-truth PDFs	40
Per-family citation PDFs	40

Scope

OEP-Bench focuses on several enterprise knowledge-system challenges:

• current and historical policy-version retrieval;
• temporal reasoning over policy effective dates;
• permission-aware retrieval and access-control scenarios;
• citation and version correctness;
• information-not-found behavior;
• prompt-injection and permission-denied cases;
• cross-document and policy-conflict reasoning;
• near-duplicate and semantic retrieval cases.

The benchmark is intended for research and evaluation of enterprise retrieval, RAG, knowledge-agent, and related information-access systems.

Release Structure

The v1.0 release is organized as follows:

OEP-Bench/
├── README.md
├── data/
│   ├── citations/
│   ├── documents/
│   │   └── addenda/
│   ├── ground_truth/
│   ├── metadata/
│   ├── organization/
│   ├── permissions/
│   ├── questions/
│   └── splits/
└── machine_readable/
    ├── cases.jsonl
    ├── documents.jsonl
    └── permissions.jsonl

The data/ directory preserves human-readable benchmark artifacts, while machine_readable/ provides normalized JSONL files for programmatic use.

machine_readable/cases.jsonl

Contains one record for each of the 884 benchmark cases.

Each record contains 14 fields:

• question_id
• family_code
• question_type
• difficulty
• question
• expected_action
• expected_answer
• supporting_evidence
• gold_document_id
• gold_version
• gold_page
• gold_section
• gold_chunk_id
• permission_expectation

The expected_action field distinguishes benchmark semantics such as answering, denying, and refusing. permission_expectation is preserved separately so that access-control expectations are not inferred solely from the expected action.

Some cases have no gold document or citation fields. These include cases where the benchmark expects behavior such as information-not-found handling or refusal of a prompt-injected instruction. Empty gold fields should therefore not automatically be treated as missing annotations.

machine_readable/documents.jsonl

Contains 121 retrievable document records:

• 119 canonical main-policy versions;
• 2 E1 jurisdiction-specific addenda.

Each record contains:

• document_id
• family_code
• document_type
• document_code
• title
• version
• status
• department
• effective_date
• classification

document_type distinguishes ordinary main_policy records from addendum records.

The 119 main-policy records are distributed by version as follows:

Version	Main-policy records
1.0	40
2.0	40
3.0	39

The E5 Whistleblower family has no version 3.0. This is intentional and should not be filled or inferred by downstream users.

machine_readable/permissions.jsonl

Contains 121 permission records, one for every retrievable document record.

Each record contains:

• family_code
• document_id
• version
• document_code
• classification
• permission_level_required
• allowed_roles
• allowed_departments
• notes

For benchmark evaluation, the explicit role and department permissions should be preserved rather than reducing access decisions to the numeric permission level alone.

Human-Readable Artifacts

The release includes human-readable benchmark artifacts corresponding to the machine-readable evaluation data:

• data/documents/ — 119 canonical main-policy PDFs;
• data/documents/addenda/ — 2 E1 addenda in Markdown;
• data/metadata/ — 40 per-family metadata PDFs;
• data/permissions/ — 40 per-family permission PDFs;
• data/questions/ — 40 per-family question PDFs;
• data/ground_truth/ — 40 per-family ground-truth PDFs;
• data/citations/ — 40 per-family citation PDFs;
• data/organization/ — the fictional NovaCore organizational setting, synthetic personnel roster, and role definitions.

Construction-time challenge-design documents, internal validation artifacts, generated reports, and quarantined artifacts are not part of the public v1.0 release.

Evaluation Case Composition

OEP-Bench contains 884 evaluation cases with unique question IDs.

Expected Actions

Expected action	Cases
ANSWER	763
DENY	40
REFUSE	81
Total	884

The expected action and the behavior of an evaluated agent should not be conflated.

In particular, a benchmark case labeled DENY can still contain substantive policy evidence describing why an action or request is not permitted. Conversely, REFUSE cases include scenarios such as information not found and prompt injection.

Permission Expectations

Permission expectation	Cases
granted_all_roles	745
granted	59
denied	80
Total	884

Permission expectation is represented independently from expected_action. This separation allows access-control behavior to be evaluated without assuming that every denied-permission case has identical answer semantics.

Gold Evidence Coverage

Of the 884 cases:

• 795 reference a gold document;
• 89 have no gold document recorded;
• all referenced gold document IDs resolve to the released document manifest;
• all populated gold versions match their corresponding released document records.

Four benchmark cases target the two E1 jurisdiction-specific addenda.

Policy-Family Distribution

Most policy families contain 22 evaluation cases. E1 Data Privacy contains 26 cases because it additionally covers the two jurisdiction-specific addenda.

Therefore:

• 39 families × 22 cases = 858 cases;
• E1 = 26 cases;
• total = 884 cases.

Historical Split Information

The release includes data/splits/split_summary.json, which preserves aggregate statistics for a historical train/dev/test split design.

The recorded aggregate counts are:

Split	Cases	Proportion
Train	508	0.5747
Dev	126	0.1425
Test	250	0.2828
Total	884	1.0000

The historical split placed all 40 prompt-injection cases in the test partition and distributed the other question types across train, dev, and test.

Important Split Limitation

The canonical per-question assignment files for the historical split are not included in this release and are not recoverable from the currently available benchmark artifacts.

split_summary.json contains aggregate counts and question-type distributions only. It does not contain question IDs or per-case split assignments.

Accordingly:

• the aggregate split statistics are provided for transparency and historical reference;
• users should not infer or reconstruct per-question assignments and describe them as the original canonical split;
• experiments using all 884 cases should explicitly report that they evaluate the full benchmark;
• researchers who require train/dev/test partitions may define a new derived split, but should document the method and random seed and identify it as a derived split rather than the original historical split.

Benchmark Semantics

Version Semantics

Policy versions are identified explicitly. Historical questions may target a policy version that was effective at the time referenced by the question rather than the latest version.

For the main policy corpus:

• version 1.0 contains 40 records;
• version 2.0 contains 40 records;
• version 3.0 contains 39 records;
• E5 intentionally has no version 3.0.

Downstream systems should preserve exact version identities rather than synthesizing missing versions.

Permission Semantics

Permission information is represented explicitly in permissions.jsonl.

The benchmark distinguishes:

• cases available to all benchmark roles;
• cases granted under specific permission conditions;
• cases where access is denied.

The explicit allowed_roles and allowed_departments fields are part of the benchmark permission model. The numeric permission_level_required field should not be treated as a complete replacement for these explicit permissions.

Refusal and Denial Semantics

ANSWER, DENY, and REFUSE are benchmark annotations with distinct meanings.

A DENY case does not necessarily mean that an evaluated system should produce an evidence-free refusal. It can require a substantive answer explaining that a policy does not permit a requested action.

REFUSE includes multiple scenario types, including information-not-found and prompt-injection cases. Evaluation systems should therefore preserve the benchmark annotations rather than mapping all non-ANSWER cases to a single generic refusal behavior.

Citation Semantics

Gold citation fields identify the expected evidence where applicable. Citation evaluation can consider document identity, policy version, section, page, and chunk information depending on which gold fields are populated.

Not every case has a populated gold citation. Missing citation fields can be intentional and should be interpreted together with the case's expected action and scenario type.

Evaluation Guidance

OEP-Bench supports evaluation across several dimensions of enterprise retrieval and RAG behavior. Researchers should report the population and denominator used for each metric because not every metric is applicable to every benchmark case.

Retrieval

Useful retrieval measures include:

• Hit@1;
• Hit@5;
• reciprocal rank;
• version accuracy at rank 1;
• whether the gold policy version appears in the retrieved evidence.

Retrieval metrics that require a gold document should be calculated only over cases for which the corresponding gold evidence is defined.

Citation Correctness

Citation evaluation can separately measure:

• correct gold document citation;
• correct gold document-version citation.

This distinction is important because retrieving or citing the correct policy family while using the wrong historical version can still produce temporally incorrect evidence.

Refusal Behavior

Refusal correctness should be evaluated over cases where refusal behavior is applicable rather than over all 884 cases.

In particular, DENY and REFUSE should remain distinct benchmark semantics. A permission-related policy answer should not automatically be scored as an agent refusal simply because the policy denies the requested action.

Permission and Security Behavior

Permission-aware evaluation should preserve the benchmark's explicit access-control annotations and report the applicable denominator.

Systems should apply authorization constraints before exposing evidence to downstream answer generation. Evaluation should not treat unauthorized evidence as acceptable merely because the retrieved text is otherwise relevant.

Temporal and Version-Aware Evaluation

For questions referring to historical periods, the applicable policy version may differ from the latest active version. Systems should therefore evaluate version correctness independently from document-family relevance.

A retrieval result from the correct policy family but the wrong effective version should not be treated as fully version-correct.

Answer Evaluation

The release contains expected answers and supporting evidence, but v1.0 does not prescribe a single semantic answer-similarity metric.

Researchers using lexical, model-based, human, or other semantic answer-evaluation methods should document those methods separately rather than presenting them as an official OEP-Bench metric.

Runtime and Cost

Latency, token usage, and generation cost may be reported as system diagnostics. Such measurements depend strongly on hardware, model configuration, caching, retrieval infrastructure, and execution conditions and should therefore be accompanied by the relevant experimental setup.

Recommended Reporting

For reproducible OEP-Bench experiments, report at minimum:

• the benchmark release/version used;
• the number of evaluated cases;
• whether evaluation used all 884 cases or a separately defined subset;
• the retrieval method;
• the generation model, if applicable;
• the permission model used during retrieval;
• the policy-version selection method;
• retrieval depth or k;
• citation-generation method;
• metric definitions and denominators;
• handling of ANSWER, DENY, and REFUSE;
• handling of cases without gold documents;
• any derived train/dev/test split and its construction method;
• relevant runtime environment and model configuration.

Results from differently defined subsets or derived splits should not be presented as directly equivalent without identifying those differences.

Synthetic Enterprise Setting

OEP-Bench is situated within NovaCore Technologies, a fictional enterprise created as the organizational setting for the benchmark.

NovaCore provides a consistent environment in which policy ownership, departments, personnel roles, reporting relationships, document classifications, and access-control scenarios can be represented across the benchmark.

The released organizational artifacts include:

• a fictional company profile;
• department charters;
• an organizational chart;
• a synthetic personnel roster containing 30 benchmark personas;
• role definitions covering 11 benchmark roles.

The personnel records are benchmark personas rather than real employees of an organization. The released roster does not include personal email addresses, telephone numbers, street addresses, dates of birth, passwords, credentials, or similar personal contact or authentication information.

Roles and Authorization

The organizational setting defines roles such as employees, managers, security personnel, auditors, executives, and administrators.

Authorization decisions should use the explicit permission information associated with each document. In particular, allowed_roles and allowed_departments carry benchmark access-control information that should not be replaced by a simple numeric role hierarchy.

This design allows OEP-Bench to represent cases in which a document may be relevant to a query but should not be exposed to the requesting benchmark identity.

Intended Use

OEP-Bench is intended for research and evaluation involving:

• enterprise information retrieval;
• retrieval-augmented generation;
• permission-aware retrieval;
• version-aware and temporal retrieval;
• citation correctness;
• refusal and denial behavior;
• secure enterprise knowledge agents;
• evaluation of retrieval strategies under a shared enterprise-RAG pipeline.

The benchmark may also be useful for developing new evaluation methods, retrieval approaches, and controlled enterprise knowledge-system experiments.

OEP-Bench is an evaluation benchmark and synthetic research environment. It should not be interpreted as a set of operational policies for a real organization, legal advice, compliance guidance, or a substitute for an organization's own authorization and information-governance controls.

Limitations

OEP-Bench is a controlled synthetic benchmark and does not reproduce the full complexity of a production enterprise environment.

Users should consider the following limitations when interpreting results:

• Synthetic organization: NovaCore Technologies is fictional. Results on OEP-Bench do not by themselves establish performance on real enterprise policy collections.
• English-language benchmark: The released policies, questions, annotations, and organizational artifacts are primarily designed for English-language evaluation.
• Single organizational setting: The benchmark models one fictional enterprise with a fixed organizational structure, policy taxonomy, and permission model.
• Controlled permission model: Authorization is represented through synthetic roles, departments, classifications, and explicit document permissions. Real enterprise authorization systems may include substantially more complex identity and policy rules.
• Historical split assignments unavailable: The release preserves aggregate statistics for a historical train/dev/test design, but the canonical per-question assignments and original random seed are not available in the current benchmark artifacts.
• No released canonical chunk corpus: The v1.0 release does not prescribe a canonical set of retrieval chunks. Researchers may therefore use different chunking strategies, but should report them as part of the experimental configuration.
• Mixed document formats: Canonical main-policy documents are released as PDFs, while the two E1 jurisdiction-specific addenda are released as Markdown.
• No prescribed semantic answer metric: Expected answers and supporting evidence are provided, but v1.0 does not define a single official semantic answer-similarity metric.
• No multi-turn benchmark component: The v1.0 release evaluates individual benchmark cases and does not include a canonical multi-turn conversation dataset.
• System-dependent runtime measurements: Latency, token usage, and related runtime measurements are dependent on the evaluated implementation and execution environment.

These limitations should be considered when comparing systems or generalizing benchmark results.

Known Release Notes

For OEP-Bench v1.0:

• the canonical main-policy corpus contains exactly 119 versions;
• the two E1 addenda are separate retrievable documents, producing 121 total retrievable document records;
• E5 intentionally contains versions 1.0 and 2.0 only;
• four evaluation cases reference the E1 addenda;
• the machine-readable case set contains exactly 884 unique question IDs;
• the release contains 121 document records and 121 corresponding permission records;
• historical aggregate split statistics are included, but canonical per-question split assignments are not;
• construction-time challenge-design documents are not part of the public v1.0 release;
• generated experiment reports, local database state, internal validation artifacts, and quarantined construction artifacts are not part of the public dataset release.

Versioning and Maintenance

OEP-Bench uses explicit release versions so that experimental results can be associated with a specific benchmark state.

This dataset represents OEP-Bench v1.0.

Researchers should record the benchmark version used in experiments and publications.

Release Stability

The contents of a published benchmark version should be treated as fixed for reproducible evaluation. Changes that alter evaluation cases, document identities, policy versions, permissions, gold evidence, or benchmark semantics should be released under a new benchmark version rather than silently modifying an existing release.

Minor documentation corrections that do not change benchmark semantics may be recorded separately in the release history.

Derived Resources

Researchers may create derived resources such as:

• alternative document chunking schemes;
• embeddings or vector indexes;
• alternative train/dev/test partitions;
• reformatted document representations;
• additional annotations;
• system-specific retrieval indexes.

Such resources should identify the OEP-Bench version from which they were derived and should not be described as canonical OEP-Bench artifacts unless they are incorporated into an official benchmark release.

Reporting Corrections

If an issue is discovered in a released benchmark artifact, it should be documented with enough information to identify:

• the affected benchmark version;
• the affected files or cases;
• whether evaluation semantics are changed;
• the correction introduced in a subsequent release.

This versioning approach is intended to preserve reproducibility across published OEP-Bench experiments.

Citation

If you use OEP-Bench in research, evaluation, or a derived resource, please cite the benchmark and identify the release version used.

A formal citation entry for the accompanying OEP-Bench research paper will be added when the paper's publication metadata is available.

Until then, please attribute the dataset as:

Pudamya Rathnayake. OEP-Bench: Orex Enterprise Policy Benchmark. Version 1.0, 2026.

Creator and Maintainer

OEP-Bench was created and is maintained by Pudamya Rathnayake, affiliated with Orex Nova.

The canonical public dataset repository is maintained under the Hugging Face account Pudamya.

NovaCore Technologies is the fictional enterprise represented within the benchmark and is not the creator, publisher, or maintainer of OEP-Bench.

License

OEP-Bench v1.0 is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Under CC BY 4.0, you may share and adapt the benchmark for any purpose, including research and commercial use, provided that appropriate attribution is given, a link to the license is provided, and changes are indicated where applicable.

The license applies to the OEP-Bench materials distributed as part of this public release. Third-party materials, if separately identified, are not relicensed by OEP-Bench.

For the full license terms, see the LICENSE file included with this release.