Use a private GitHub Security Advisory for suspected credential exposure, dependency vulnerabilities, unsafe document handling, or prompt-injection paths. Do not disclose exploitable details in a public issue before a fix is available.

• Credentials are loaded from the process environment or an ignored .env file.
• .env.example contains placeholders only.
• Logs and reports must never include authorization headers or credential values.
• Rotate a credential immediately if it is committed, pasted into an issue, or exposed in terminal output.

The publish-readiness script scans maintained text files for common provider-key formats. This is a guardrail, not a replacement for repository secret scanning.

Hosted analysis sends relevant contract text to the configured model provider. Do not process privileged, confidential, personal, export-controlled, or otherwise restricted material unless the provider, account configuration, and organizational policy explicitly permit it.

The deterministic test and benchmark workflows operate on bundled public or synthetic fixtures. Generated reports are ignored by default because they may contain source contract language.

Contracts are treated as untrusted data rather than instructions. Model prompts separate system policy from serialized document content, and model output is validated before entering application state. Consumers should still review output for prompt injection, fabricated authority, and adversarial formatting.

The browser control plane limits request size, sanitizes uploaded names, stores files under per-job directories, and rejects unsupported extensions. The Python loader separately bounds source size, DOCX expansion and entry count, PDF page count, and extracted text length before analysis.

API responses are marked no-store; browser assets use a restrictive Content Security Policy and same-origin resource policy. The server binds to loopback by default. Deployments that bind to a non-loopback interface require authentication, TLS termination, access logging, retention controls, and an explicit network trust boundary in front of ClauseGuard.

ClauseGuard is a review aid, not an authorization or enforcement system. Do not automatically execute contractual decisions from a finding or rewrite. Preserve the source document, JSON audit record, model configuration, and human approval when findings influence a legal workflow.

CI installs dependencies from pyproject.toml, runs static analysis and tests, and exercises parsers with malformed inputs. Review dependency updates before merging, especially document parsers and network clients.

CI audits Python, npm, and Go dependency graphs and runs CodeQL across Python, Go, and TypeScript. Dependabot monitors all four dependency ecosystems, including GitHub Actions.