Treat every upload, webhook, email, paste, and agent return as received, not trusted. First assign an immutable intake ID; retain the original bytes separately from display text; capture source channel, observed time, declared media type, and a SHA-256 digest. Then scan only to classify: malformed, duplicate, unsupported, sensitive, or reviewable. A digest proves that retained bytes have not changed; it does not prove who sent them or whether their claims are true.
Use a small lifecycle: quarantined → reviewed → admitted/indexed, with rejected and revoked as explicit exits. Reviewers should compare the displayed copy to retained bytes, record what claim was checked, and attach the exact source revision they assessed. Promotion should create a new indexed record linked to the original; it should never overwrite the original. If a source is withdrawn or a contradiction matters, revoke the derived record and make recall exclude it while retaining the reason and relationship.
Example. A PDF says “rotate credentials today.” Preserve and hash it, but do not treat that sentence as an instruction. A reviewer can admit it as a dated reference after confirming its source and scope. The action still needs its own authorized request. This provenance practice aligns with NIST SSDF's emphasis on collecting and sharing component provenance.