Regula

Offline-capable, code-native AI governance scanning. The core scan runs locally, does not upload scanned file contents, and needs no account; telemetry is sent only with explicit opt-in consent. Regula flags patterns that may need review under the EU AI Act, South Korea's AI Basic Act, and Colorado SB 26-189, records the deployment facts code cannot show, and reports insufficient information rather than inventing a score.

Table of contents

• What it does
• Choose how to start
• Quick start
• What Regula tells you
• Key commands
• Who is this for?
• What Regula is (and isn't)
• Bias evaluation: methodology and ethics
• Important limitations
• Verified numbers
• Contributing
• Licence

$ regula check examples/cv-screening-app --scope all

Decision: insufficient_information
Jurisdiction: eu
Rule resolution: unresolved
Facts needed to resolve the next decision: 2
  - is_ai_system: Does the subject meet the governing law's definition of an AI system or regulated automated technology?
  - jurisdiction_in_scope: Does this jurisdiction's territorial and operator scope apply?

Detector observations (not legal facts):

  Detector summary: ANNEX III OR SECURITY PATTERNS
  The scanner found patterns relevant to Annex III or security review.
  Resolve the facts listed above before attaching Article 9 to 15 duties.
  Files scanned:      1
  High-risk:          1
  INFO tier:          1

  HIGH-RISK INDICATORS:
    [INFO] [ 43] app.py — Employment and workers management [plan]

  Detector priority: 0-100 (higher = more code patterns matched; not a correctness probability)

Excerpt from the real output of the command shown, against the tracked fixture examples/cv-screening-app. The per-category counts that read zero and the next-steps footer are omitted for length; nothing else is edited. scripts/verify_transcripts.py re-runs this command on every check and fails if any line above stops appearing in its output.

What it does

Regula scans a local source-code folder for patterns that may need AI governance review. A questionnaire records facts that code cannot show, such as where and how a system will be used. Results identify candidate risk categories and link to provisions that may be relevant. Regula does not determine legal classification, compliance, or the obligations that apply to a real deployment.

Reference material covers the EU AI Act, South Korea's AI Basic Act (Act No. 20676), and Colorado SB 26-189. The core install has no required third-party runtime dependencies. Optional extras add dependencies, and some optional commands or configured features can contact external services. Assess territorial scope and data-processing duties independently.

In plain English: give Regula a source-code folder and answer questions about how the system will be used. It returns possible issues and a review trail so a developer, governance lead, or adviser can decide what to investigate next. A scan with no findings does not prove that a system is compliant or low risk.

Choose how to start

Quick start

pipx install git+https://github.com/kuzivaai/getregula.git@main

PyPI distribution is currently unavailable. The command above installs a moving public source reference; pin a reviewed commit hash for reproducible use. See docs/installation.md for the verified status and alternatives.

Not sure if the AI Act applies? No code needed:

regula assess               # record declared context for human review
regula assess --save-facts  # and write the answers where `regula check` reads them

The scan tells you which facts it needs and cannot get from code. You supply them, and the decision moves:

regula check . --list-facts                     # every fact id the model defines
regula check . --fact is_ai_system=yes \
               --fact jurisdiction_in_scope=yes # declare them for one run

Declared facts are yours, not Regula's. Each is stored with who declared it, through which command, in answer to which question, and when, in .regula/facts.json, and the scan prints that provenance beside the decision. unknown is an answer and is never read as no. A declared fact can move a decision from insufficient_information to an indication; it does not produce a risk tier, a compliance score, a readiness percentage or an effort estimate.

Want to scan your code?

regula check .              # 423 tier patterns, 8 language families; runtime varies
regula check . --jurisdictions eu,korea,colorado  # selected reference mappings

Need a review pack?

regula evidence-pack --project .   # reviewer-completable evidence scaffold
regula conform --sign --timestamp  # integrity metadata (requires regula[signing])

Generated files are inputs to human review, not an audit opinion, certification, or proof of compliance.

Just want to see it work? (requires the cloned repo : examples/ is not bundled in the pip package)

git clone https://github.com/kuzivaai/getregula && cd getregula
regula demo                 # scan a bundled example project

Install details

The recommended install is pipx : it isolates Regula from your system Python and avoids the externally-managed-environment error on Ubuntu 22.04+, Debian 12+, Fedora, Arch, and Homebrew Python.

If you don't have pipx yet, install it first (one-time):

Already using uv? uvx --from git+https://github.com/kuzivaai/getregula.git@main regula runs the public source without a persistent install. Or install it with uv tool install git+https://github.com/kuzivaai/getregula.git@main.

Running inside a venv or conda env? Use pip install git+https://github.com/kuzivaai/getregula.git@main. The PEP 668 restriction applies to managed system Python, not an activated virtual environment.

See docs/installation.md for troubleshooting (externally-managed-environment, command not found: regula after install, PATH setup per shell).

Try it against a known high-risk fixture:

regula check examples/cv-screening-app --scope all

The --scope all flag is needed because Regula's default scope (production) skips example directories. This fixture intentionally contains employment-related code indicators that exercise the Annex III Category 4 detector; the resulting label is a detector observation, not a legal determination.

See examples/ for runnable reference projects covering prohibited-practice, high-risk, transparency, and no-elevated-indicator scenarios, or walk through the full 10-minute evaluation journey in examples/cv-screening-app/ : install, scan, plan, gap, conform, verify, handoff to red-team tooling.

For a deeper first-time-user walk-through (policy tuning, CI integration, baselining) see docs/QUICKSTART.md. The full documentation is indexed by type (tutorials / how-to / reference / explanation) in docs/README.md.

CI/CD

# .github/workflows/regula.yaml
name: AI Governance Check
on: [push, pull_request]
jobs:
  regula:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: kuzivaai/getregula@v1
        with:
          path: '.'
          upload-sarif: 'true'
          fail-on-prohibited: 'true'

What Regula tells you

The EU AI Act defines four risk tiers. Regula maps code patterns to each:

Every finding includes the relevant Article reference and explains when exceptions may apply. Regula flags patterns -- it does not make legal determinations.

Coverage has explicit levels. Regula has an evidence-gated questionnaire and decision model for the EU AI Act, South Korea's AI Basic Act (Act No. 20676), and Colorado SB 26-189. It separately maps findings to selected control or provision references in 13 frameworks, and publishes dated regional trackers. A crosswalk or tracker is not an executable applicability decision or a compliance assessment. See product coverage and user journeys for the current capability matrix and claim boundary. Use --jurisdictions on regula check to filter selected reference mappings; that option also accepts uk, brazil, nist, and iso, but it does not run an applicability decision. Use --jurisdiction eu, korea, or colorado on regula assess for the three implemented decision-support configurations.

Developer guide sources: Python | JavaScript | Healthcare | Recruitment | Article 5 | Article 9 | Article 14 | Article 50

Key commands

Regula has 62 commands in total. Run regula --help-all for the full list, or see docs/cli-reference.md.

REST API and web dashboard

For GRC integration or non-terminal users:

python3 scripts/api_server.py --port 8487
# Open http://localhost:8487/v1/dashboard

Seven endpoints: /health, /v1/check, /v1/classify, /v1/gap, /v1/questionnaire, /v1/questionnaire/evaluate, /v1/dashboard. All return the same JSON envelope as the CLI. No auth -- run behind a reverse proxy for remote access.

Who is this for?

• Builders and maintainers who need an initial list of source-code signals to investigate before contextual and legal review.
• Governance and assurance reviewers who need observations, declared facts, unresolved questions, and provenance kept visibly separate.
• Evaluators and adopters who need reproducible tests, completion data, known failure modes, and a non-blocking path to test Regula on their own representative sample.
• Qualified contextual reviewers who may use Regula artefacts as inputs. Regula does not make their legal, conformity, security, accessibility, or domain determination.

The detailed tasks, failure paths, and required interface states are defined in product coverage and user journeys.

What Regula is (and isn't)

Regula is:

• A development-time evidence and triage tool that combines static code analysis with governance questionnaires, mapping observations and declared facts to possible obligations across 3 jurisdictions (EU AI Act, South Korea AI Basic Act, Colorado SB 26-189)
• A shift-left code-indicator scanner -- like ESLint for governance review, running in your terminal or CI/CD pipeline
• A questionnaire-based assessment tool for organisational obligations that code patterns cannot verify (Articles 9, 17, 27, 72)
• Pattern-based risk indication across 3 jurisdictions, not a legal compliance certificate
• A starting point for compliance awareness, not a finish line

Regula is not:

• A runtime monitoring system (it analyses source code, not running systems)
• A legal compliance certificate (findings are indicators, not legal determinations)
• A replacement for enterprise GRC platforms like Credo AI or Holistic AI (it complements them)
• A production fairness testing platform (regula bias runs benchmark probes against a local model as a starting point, but does not replace runtime fairness monitoring)
• Legal advice (consult qualified legal counsel for compliance decisions)

Regula helps development teams find AI-governance questions earlier across its implemented jurisdictions and reference frameworks. It does not replace the organisational, procedural, domain, and legal work required for compliance. For a detailed account of what falls outside Regula's scope, see docs/what-regula-does-not-do.md, and for Regula's own model card (intended use, training data, evaluation, known failure modes) see docs/MODEL_CARD.md.

Bias evaluation: methodology and ethics

regula bias runs two social-bias benchmarks against a locally-hosted language model (Ollama, llama3.2/mistral/qwen variants supported) as evidence for EU AI Act Article 10 data-governance documentation.

Both include Wilson confidence intervals for small-sample reliability and bootstrap CIs for distribution estimates. Full methodology lives in scripts/bias_eval.py and docs/benchmarks/PRECISION_RECALL_2026_04.md.

Ethics statement. CrowS-Pairs and BBQ stereotype pairs are used solely for scientific evaluation of model behaviour under controlled conditions. Regula does not display individual stereotype pairs in terminal output or reports : only aggregated scores, confidence intervals, and benchmark-level verdicts. The pairs are distributed under the dataset's own licence (CC BY-SA 4.0 for CrowS-Pairs) and are not redistributed or modified by Regula. Opinions encoded in the stereotype pairs do not reflect the views of the maintainer, Regula contributors, or any user running the tool; their presence is instrumental, not endorsing. regula bias is a development-time starting point for bias documentation, not a production fairness monitor : see "What Regula is (and isn't)" above.

Important limitations

Regula performs pattern-based risk indication, not legal risk classification.

• The EU AI Act classifies risk based on intended purpose and deployment context (Article 6), not code patterns. Regula's findings are indicators that warrant human review.
• False positives will occur. Regula has no current, independently labelled real-world precision estimate. Its development corpora are single-reviewer records, and the older random-corpus result cannot be re-derived from a clean checkout because its measured subset and pinned source snapshots were not preserved. The arithmetic, limitations, and superseded measurement remain available as a dated research record in benchmarks/README.md; they are not evidence of the current detector's accuracy.
• TypeScript findings are advisory: current real-world precision is unmeasured. A dated six-finding development slice contained six false positives, which is a warning signal rather than a dependable rate.
• False negatives will occur. Novel risk patterns not in the database will be missed.
• Article 5 prohibitions have conditions and exceptions that require human judgment.
• The audit trail is self-attesting (locally verifiable, not externally witnessed).
• This is not a substitute for legal advice or DPO review.

Verified numbers

For reproduction commands, version-bounded benchmarks, known exceptions, security posture, and audit-trail design, see docs/TRUST.md. What version numbers promise, the public API they cover, and the deprecation policy: docs/VERSIONING.md.

Privacy and data handling

Regula's core scan runs on your machine. No code, findings, or metadata are transmitted during a default scan. There is no account system or API key, and no telemetry by default (crash reporting requires both regula telemetry enable and an endpoint you configure yourself via REGULA_SENTRY_DSN; published builds ship none, and DO_NOT_TRACK suppresses it regardless). The tool reads your source files, analyses them locally, and writes output to your local filesystem. Explicit network-enabled operations include RFC 3161 timestamping, feed or update checks, and integrations you configure. See SECURITY.md for the full security posture.

Contributing

Bug reports and pull requests are welcome.

• Run pytest tests/ -q before opening a PR.
• Pattern additions go in scripts/risk_patterns.py. Each pattern should have a corresponding test.
• Regula is intentionally risk indication, not legal classification. New patterns should be conservative -- false positives erode trust more than false negatives for a developer tool.
• See CONTRIBUTING.md for the full contributor guide and CHANGELOG.md for version history.

Authorship

Regula currently has a single maintainer. Where commits identify a co-author, that attribution records the tools or collaborators involved. The maintainer remains accountable for reviewing and accepting every merged change. See GOVERNANCE.md for the bus-factor limitation.

Licence

Engine and CLI: Apache License 2.0 OR European Union Public Licence v. 1.2 : at your option. Pick the one that fits your context:

• Apache 2.0 includes an explicit patent grant, making it the preferred choice for enterprise adoption, commercial redistribution, and any context where patent clarity matters.
• EUPL-1.2 is explicitly recognised inside EU institutions and public-sector procurement, is strongly-copyleft on software, and has a formal compatibility appendix (GPL v2/v3, AGPL v3, OSL, EPL, CeCILL, MPL 2.0, LGPL, CC BY-SA 3.0) for downstream projects. If you work with a European public administration, EUPL is often the required or preferred licence.

You may choose either licence for any use. You do not need to state which one you picked, but attribution (keep the copyright notice and NOTICE file) is required under both.

Risk patterns and regulatory data: Detection Rule License (DRL) 1.1. You may use, modify, and redistribute the patterns freely. Attribution is required if you redistribute the patterns or use them in a product. If your tool generates match output from these patterns, the output must credit the source.

The SPDX expression for the full package is (Apache-2.0 OR EUPL-1.2) AND LicenseRef-DRL-1.1.