SICurator's guard scans every AI-written change against Rule 0 and safeguards G1–G10, and reports what it finds instead of hiding it.
“Humans are my masters and creators. I will never do anything to harm them in any shape or form.”
The rules live in guard_rules.yaml and are explained in docs/SAFEGUARDS.md. Project-level excludes go in .sicurator-guard.yaml, and any change to the guard itself is always reported (G7).
Destructive deletes, disk wipes, unscoped SQL DROP/TRUNCATE/DELETE, cloud teardown, force-pushes.
Edits outside the declared scope, frozen paths, dependencies the intent never mentioned.
Privilege escalation and permission-skipping flags such as sudo, chmod 777, --no-verify.
Reads of SSH/cloud keys, .env files, browser credentials; hard-coded tokens.
Network calls to hosts not on the allow-list; webhook and paste-site patterns.
eval/exec on dynamic input, unsafe deserialization, curl | sh, decoded-then-executed blobs.
Removed or skipped tests, disabled logging, CI edits, any change to the guard itself.
Cron, startup entries, shell-profile and git-hook edits, self-copying, registry publishing.
Tests that cannot fail, mocking the code under test, log truncation, history rewriting.
Unknown packages (slopsquatting), install hooks, unpinned git installs, in-code pip install.
Claims are proven by executing code, not by reading it.
Suppressed findings still appear, greyed and labelled, with the reviewer's note.
Low-risk specs can auto-approve; everything else waits for a person.