Glassbox EU AI Act Compliance Audit

Version updated for https://github.com/designer-coderajay/Glassbox-AI-2.0-Mechanistic-Interpretability-tool to version v3.0.0.

  • This action is used across all versions by ? repositories.

Action Type

This is a Composite action.

Go to the GitHub Marketplace to find the latest changes.

Action Summary

Glassbox 3.0.0 is an open-source toolkit designed to automate the creation of Annex IV technical documentation required under the EU AI Act for high-risk AI systems, streamlining compliance processes for both open-source and proprietary AI models. It also provides advanced tools for researchers to identify and analyze the minimum faithful circuit in transformer models, offering significant performance improvements in mechanistic interpretability tasks. Key capabilities include automated compliance report generation, black-box model auditing, and a REST API for integration into workflows.

Release notes

Glassbox v3.0.0 — Enterprise EU AI Act Compliance Platform

Enforcement deadline: August 2026. Every high-risk AI system in financial services, healthcare, HR, and legal must be able to explain its decisions. Glassbox is the audit infrastructure.

What’s new in v3.0.0

🧪 BiasAnalyzer — EU AI Act Article 10(2)(f)

Three bias tests built for regulatory submission:

  • counterfactual_fairness_test() — swaps demographic attributes in prompt templates, measures probability gap across groups
  • demographic_parity_test() — flags positive outcome rate disparity across demographic groups
  • token_bias_probe() — detects stereotypical associations between demographic and role tokens
  • Works offline (pre-computed logprobs) or online (live model_fn)
  • BiasReport.to_markdown() generates Annex IV Section 5 ready for submission
from glassbox import BiasAnalyzer
ba = BiasAnalyzer()
result = ba.counterfactual_fairness_test(
    prompt_template="The {attribute} applied for the loan",
    groups={"gender": ["male applicant", "female applicant"]},
    target_tokens=["approved", "denied"],
    model_fn=my_model,
)

🔔 Webhooks

Full webhook system on the REST API:

  • POST /v1/webhooks — register callback URL with HMAC-SHA256 signing
  • Fires on job.completed and job.failed events
  • X-Glassbox-Event + X-Glassbox-Signature headers on every delivery

📊 Multi-Audit History Dashboard

  • F1 faithfulness trend chart (Chart.js) with Grade C threshold line
  • Grade distribution bar chart (A/B/C/D counts)
  • Audit table with Date/Model/Grade/F1/Status
  • “Load from API” button connects to GET /v1/audit/reports

🖼️ Circuit SVG Export

  • “Download SVG” button in the D3 circuit graph panel
  • Exports standalone glassbox-circuit.svg with inlined dark-mode styles
  • Paper-ready at any resolution

📋 Risk Register — Article 9

Persistent cross-audit risk tracking:

  • ingest_annex_report() auto-extracts Section 5 risks from any AnnexIVReport
  • Deduplication + occurrence counting across sessions
  • Status lifecycle: open → mitigated | accepted | escalated
  • trend_summary() returns compliance_health: green | amber | red
  • to_markdown() for embedding in PRs and reports
from glassbox import RiskRegister
rr = RiskRegister("risks.json")
rr.ingest_annex_report(annex, model_name="gpt2")
print(rr.trend_summary())  # {'compliance_health': 'amber', 'open': 2, ...}

🧪 Test Suites

  • 76 new tests: tests/test_audit_log.py (32) + tests/test_widget.py (44)
  • All offline with tmp_path fixtures

Install

pip install glassbox-mech-interp==3.0.0

Quick start

from glassbox import GlassboxV2, AuditLog, BiasAnalyzer, RiskRegister, AnnexIVReport

gb   = GlassboxV2(model)
log  = AuditLog("audit.jsonl")
rr   = RiskRegister("risks.json")

result = gb.analyze("The capital of France is", " Paris", " London")
log.append_from_result(result)

annex = AnnexIVReport(model_name="GPT-2", provider_name="Acme Corp",
                       provider_address="London, UK",
                       system_purpose="Financial decision support",
                       deployment_context=DeploymentContext.FINANCIAL_SERVICES)
annex.add_analysis(result)
rr.ingest_annex_report(annex)

Full changelog: CHANGELOG.md
PyPI: pip install glassbox-mech-interp
HuggingFace Space: designer-coderajay/Glassbox-AI-2.0-Mechanistic-Interpretability-tool

Full Changelog: https://github.com/designer-coderajay/Glassbox-AI-2.0-Mechanistic-Interpretability-tool/compare/v3.0.0...v3.0.0