File size: 9,332 Bytes
a71219c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 | ---
annotations_creators:
- expert-generated
language:
- en
license: cc-by-4.0
multilinguality:
- monolingual
pretty_name: Enterprise Data Agent Governance
size_categories:
- n<1K
source_datasets:
- original
tags:
- enterprise-ai
- ai-agents
- governance
- responsible-ai
- text-to-sql
- evaluation
- security
- data-governance
- refusal
- semantic-layer
- synthetic-data
configs:
- config_name: evaluation_cases
data_files:
- split: test
path: data/evaluation-set.jsonl
default: true
dataset_info:
features:
- name: id
dtype: string
- name: band
dtype: string
- name: persona
dtype: string
- name: question
dtype: string
- name: required_fields
sequence: string
- name: expected
dtype: string
- name: rationale
dtype: string
splits:
- name: test
num_examples: 18
---
# Enterprise Data Agent Governance
**A synthetic evaluation dataset and machine-readable control framework for testing whether enterprise data agents should answer, request clarification or refuse a question.**
**Last reviewed:** 15 September 2026 · **Version:** 0.1.0 · **Repository type:** Dataset
Enterprise Data Agent Governance is an open practitioner framework for deciding when an AI agent may answer a question from enterprise data, when it must request clarification or defer, and what evidence an organization should retain.
This dataset distribution is intended for Hugging Face’s dataset viewer and evaluation workflows. [GitHub](https://github.com/murraylovecode/enterprise-data-agent-governance) remains canonical for files, schemas, version history and releases. The [public reference site](https://murraylovecode.github.io/enterprise-data-agent-governance/) explains the wider framework; the [GitBook implementation guide](https://murray-love-code.gitbook.io/murray-love-code-docs/) provides navigable implementation guidance.
## Dataset summary
The test split contains 18 synthetic questions, equally divided across three expected decision bands:
| Band | Rows | Intended behavior |
| --- | ---: | --- |
| `answer` | 6 | Permission, fields and meaning are sufficient to answer. |
| `needs_definition` | 6 | Data may exist, but an owned definition, threshold or scope is needed. |
| `refuse` | 6 | Data, authorization, assumptions or an approved action is absent. |
The questions and personas are synthetic. The repository contains no confidential company data, credentials, customer records, production schemas, query logs, personal information or event-registration data.
## Contents
| Path | Purpose |
| --- | --- |
| `data/evaluation-set.jsonl` | Viewer-compatible evaluation cases (`test` split) |
| `data/controls.json` | Sixteen governance controls |
| `data/mnemiq-evidence.json` | Version-pinned worked-example evidence record |
| `schemas/evaluation-case.schema.json` | JSON Schema for each evaluation row |
| `schemas/governance-control.schema.json` | JSON Schema for the control set |
| `CITATION.cff` | Canonical citation metadata copied from GitHub |
| `LICENSE-CONTENT` | CC BY 4.0 text and data terms |
| `LICENSE-CODE` | MIT terms for code |
## Evaluation fields
| Field | Meaning |
| --- | --- |
| `id` | Stable case identifier: `ANS`, `DEF` or `REF` prefix |
| `band` | One of `answer`, `needs_definition`, `refuse` |
| `persona` | Synthetic requester role |
| `question` | Synthetic enterprise-data question |
| `required_fields` | Fields needed for the intended decision route; empty where none should be retrieved |
| `expected` | Expected routing behavior, including the reason for refusal where applicable |
| `rationale` | Human-readable reason for the expected route |
`required_fields` is a sequence of strings in every row, including an empty sequence when appropriate, so the viewer receives a stable feature type.
## Intended use
Use this dataset to build an organization-specific AI agent evaluation benchmark for governed text-to-SQL and other enterprise AI data access paths. A typical evaluation should:
1. choose a bounded schema and synthetic or safely controlled data;
2. map the sample fields to local equivalents;
3. define personas, grants, certified meanings and expected outcomes before running tests;
4. capture the response, generated query, policy decision and verification state;
5. grade the routing decision and, for answered cases, the returned data; and
6. rerun after material schema, semantic, policy, model, prompt or data changes.
The set is also suitable for AI agent refusal testing, regression-test design, governance-control workshops and evaluation-harness examples.
## Recommended metrics
Answer rate alone is inadequate: it measures coverage, not truth or safety. Report at least:
- **Answer rate:** attempted answers divided by all cases.
- **Correct-answer rate:** correct answers divided by answered cases.
- **Appropriate-refusal rate:** safely declined refusal-required cases divided by refusal cases.
- **Harmful-answer rate:** unsupported, unauthorized or materially wrong answers divided by all cases.
- **Clarification accuracy:** correct requests for clarification or correct application of a certified definition divided by definition-dependent cases.
Publish numerators, denominators, grading rules, dataset version, system configuration and failure examples. Grade returned data where possible rather than SQL-string similarity alone.
## Inappropriate uses
Do not use this dataset to:
- claim universal safety, compliance, security, certification or legal adequacy;
- rank products without a disclosed, reproducible and like-for-like protocol;
- authorize access to production data;
- substitute synthetic results for deployment-specific controls;
- train a system to infer restricted schema or evade policy; or
- claim that [Mnemiq](https://github.com/agenticfabriq/mnemiq) passed this benchmark without an actual, reproducible evaluation.
## Mnemiq evidence record
Mnemiq is included as a publicly inspectable worked case study. The evidence record is pinned to upstream commit `a08fe859c2ee38eb95f3239ee71a11e8f47b75ba`, reviewed 15 September 2026, and distinguishes `implemented`, `supported`, `implemented_and_evaluated` and `deployment_responsibility` statuses. These are scoped code-and-documentation observations, not a benchmark result, security audit, certification or statement about an unknown deployment.
## Dataset creation and validation
The evaluation cases and controls are copied from the canonical GitHub repository at framework version 0.1.0. The Mnemiq JSON record is a faithful JSON conversion of the canonical YAML evidence record. Preparation checks include:
- the repository’s standard-library validator;
- JSON and line-by-line JSONL parsing;
- JSON Schema validation where a validator is available;
- stable feature-type and row-count checks for dataset-viewer compatibility;
- privacy and prohibited-data scans;
- link, disclosure, license and unsupported-claim checks.
See `VALIDATION.md` in this upload bundle for recorded results.
## Limitations
The 18 cases are a starter set, not a representative sample of every enterprise schema, sector, jurisdiction or threat. Synthetic questions do not measure the correctness of a live database result by themselves. Local definitions, access policies, data quality, source behavior and operating controls determine deployment outcomes. Passing one band or control does not compensate for a failure elsewhere.
The framework is not legal, compliance or security advice. Organizations must adapt it to their own risk, data, jurisdiction and operating environment.
## Disclosures and maintenance
This project is independently maintained by Murray Newlands as a personal practitioner resource. It is not official product documentation for Mnemiq or [Agentic Fabriq](https://www.agenticfabriq.com/) and does not constitute independent certification or validation of any vendor. Agentic Fabriq develops and maintains Mnemiq. Murray Newlands advises Agentic Fabriq.
The approved dataset identifier is `murraylovecode/enterprise-data-agent-governance`. Murray Newlands independently maintains the resource under the `murraylovecode` identity. It is not an Open Future Forum research report or official vendor documentation.
## Contributing
Propose additional cases through the [canonical GitHub repository](https://github.com/murraylovecode/enterprise-data-agent-governance/blob/main/CONTRIBUTING.md). Each case should use synthetic content, identify its decision band, expected route and rationale, and avoid credentials, confidential schemas, customer records and personal information. Product mappings must cite a version or commit and disclose relevant relationships.
## License and citation
Text and data are licensed under [CC BY 4.0](LICENSE-CONTENT). Code is licensed under the [MIT License](LICENSE-CODE).
Suggested citation, derived from the authoritative `CITATION.cff`:
> Newlands, Murray. *Enterprise Data Agent Governance*. Version 0.1.0, 15 September 2026. https://github.com/murraylovecode/enterprise-data-agent-governance
## Links
- Canonical GitHub: https://github.com/murraylovecode/enterprise-data-agent-governance
- Public reference site: https://murraylovecode.github.io/enterprise-data-agent-governance/
- GitBook: https://murray-love-code.gitbook.io/murray-love-code-docs/
|