RuleProbe is a public beta CLI for measuring repository AI instruction compliance. It creates disposable sandboxes and executes only bounded provider actions, but real provider runs still involve model-generated file and command requests.

• Prefer mock and dry-run for CI and demos.
• Use API providers only with repositories and credentials you are comfortable testing.
• Keep API keys in environment variables or .env; do not commit them.
• Review .ruleprobe/ artifacts before sharing logs. Extractor raw responses are sanitized for common API key formats, but logs can still contain repository-sensitive text.
• Treat reports as measurement artifacts, not as a security control or policy enforcement guarantee.

RuleProbe blocks:

• absolute paths and path traversal
• writes to .git, .ruleprobe, and node_modules
• dangerous shell commands such as shell downloads, privilege changes, and destructive filesystem commands
• long-running allowed commands via action timeouts

These protections reduce risk; they do not make arbitrary agent execution perfectly safe.

• Code-pattern assertions are evaluated from actual changed sandbox file contents, not from provider prose alone.
• File-change assertions are evaluated against the extracted path or glob pattern where possible.
• Provider failures and missing credentials should appear as explicit failures or skipped results, not as fake compliance.
• Reports include Known Limitations so consumers can distinguish mock, dry-run, fallback extraction, skipped execution, and provider failures.
• Report snippets are short evidence excerpts, not complete source archives; review artifacts before sharing them publicly.
• Fixture benchmarks are a regression gate for extraction behavior. They are not proof of universal semantic correctness.

Open an issue with:

• RuleProbe version or commit
• provider and extractor mode
• command used
• sanitized report/debug output
• expected versus actual behavior

Do not include API keys, private repository contents, or provider secrets.