Skip to content

Security: endtree-FDE/evidencegate-ocr

Security

SECURITY.md

Security policy and evaluation boundary

EvidenceGate OCR treats model output as untrusted input. It is an evaluation and human-routing layer, not an approval or payment system.

Data

  • Use synthetic or irreversibly redacted documents in public tests.
  • Do not commit employee data, tax identifiers, bank accounts, ERP records, access tokens or API keys.
  • Keep raw production documents local unless an authorized data-processing agreement and retention policy exist.

Credentials

  • Load DASHSCOPE_API_KEY from the process or user environment.
  • Never place credentials in prompts, command arguments stored in logs, JSON evidence or screenshots.
  • Rotate a credential immediately if a secret scan or review finds exposure.

Model-output boundary

  • Preserve the raw response and exit code.
  • Reject undeclared fields and invalid evidence metadata.
  • Route visible instructions, uncertainty, obstruction and business conflicts to human review.
  • Keep unavailable locator and confidence evidence as null.
  • Mark business-critical fields with "locator_required": true when source traceability is mandatory; missing locators must route to human review.
  • Never interpret ACCEPT_CANDIDATE as business approval.

Reporting vulnerabilities

Open a private GitHub security advisory when possible. Do not attach real sensitive documents or active credentials. Include a synthetic reproducer, affected version and observed output.

There aren't any published security advisories