Candidate identity, employment and education history, authorization, preferences, application answers, documents, rankings, browser state, interview notes, and outcomes are sensitive. Normal runtime stores them only in ignored local files and SQLite.

Network access is limited to public job retrieval, explicitly requested public research, ATS application pages opened for the user, and localhost Ollama. The application has no telemetry and no cloud-LLM fallback. Source adapters use timeouts, caching, conservative request rates, and clear identification.

Job descriptions and web pages are data, not instructions. Content from them cannot change tool policy, request unrelated files, trigger outbound requests, or become a candidate fact. URLs embedded in a job description are not fetched automatically.

Browser automation will use a visible session, ordinary field filling, and explicit uncertainty flags. It will not collect credentials, expose cookies, defeat authentication, evade throttling, solve CAPTCHAs, or press the final submission control. Unsupported sites fall back to opening the page for manual use.

Generated candidate claims must resolve to one or more stable source identifiers. A fact checker compares claims with source data and rejects unsupported dates, organizations, technologies, metrics, credentials, authorization statements, and accomplishments. Gaps remain gaps.

Private runtime paths are ignored by Git. Tests use a clearly fictional profile. Logs contain identifiers, durations, status codes, and validation summaries—not profile content, prompt bodies, cookies, or answers. Never commit credentials, session data, production databases, generated applications, or real profiles.

The local machine and the user's OS account are trusted. LocalJobAgent is not a sandbox for hostile local users or malware. It does not attempt to automate sites that prohibit or technically prevent responsible access.