designer-coderajay/Glassbox EU AI Act Compliance Audit

EU AI Act Annex IV compliance gate for CI/CD. Fails the build if model explainability grade falls below your threshold.

View on GitHub

Trust Signals

Scorecard Score
not yet scored
Maintenance Recency
Activelast commit Jun 28, 2026
License
MIT

Pinned Snippet

workflow.ymlSHA-pinned
uses: designer-coderajay/glassbox-mech@3942b5cb81096f03b55fe25cae63d67357fef3ea # v4.4.0

tags can be moved; commit SHAs can't. why a SHA?

namedescriptionrequireddefault
api_urlGlassbox API endpoint. Use the public hosted API or your self-hosted instance.nohttps://designer-coderajay-glassbox-ai-2-0-mechanistic-interpretability-tool.hf.space
model_nameModel to audit (TransformerLens-compatible for white-box, e.g. gpt2, gpt2-medium)yes
promptDecision prompt for attribution analysis (the input your model makes decisions on)yes
correct_tokenExpected correct output token (e.g. " Mary", " approved")yes
incorrect_tokenDistractor / incorrect token (e.g. " John", " denied")yes
provider_nameProvider name for Annex IV Section 1 (your company/team name)noCI Pipeline
provider_addressProvider address for Annex IV Section 1no""
system_purposePurpose of the AI system (for Annex IV documentation)noAutomated decision-making system under EU AI Act review
deployment_contextDeployment context. One of: financial_services | healthcare | hr_employment | legal | critical_infrastructure | education | other_high_risk noother_high_risk
methodAttribution method. "taylor" (fast, default) or "integrated_gradients" (more accurate)notaylor
fail_below_gradeFail the CI step if the explainability grade is strictly below this. A = best, D = worst. Set to "D" to never fail on grade. noC
output_pathFile path to write the JSON compliance reportnoglassbox-report.json
namedescription
gradeExplainability grade: A (fully explainable) → D (not explainable)
f1_scoreFaithfulness F1 score (0.0–1.0)
compliance_statusCompliance status: conditionally_compliant | incomplete | non_compliant
report_idGlassbox report ID — use to download PDF/JSON from the API
sufficiencyFaithfulness sufficiency score
comprehensivenessFaithfulness comprehensiveness score