Data-Poisoning Checklist
Goal
Make sure a knowledge source is trustworthy before it can influence anything. Poisoning is the risk that a source’s content has been altered, fabricated, or silently changed so that governed decisions inherit bad input. The defense is provenance, validation, version control, change history, and rollback.
Apply this checklist when a source is registered, re-indexed, or updated — not on every read. It complements the prompt-injection-in-sources-checklist: injection is about content acting as instruction; poisoning is about content being untrustworthy in the first place.
The checklist
1. Provenance is recorded
- The source has an owner (person or team), not just an uploader.
- Origin is documented: where the content came from and who authored it.
-
source_integrity_notes(or equivalent provenance field) are present and specific enough to be reviewable. - For external sources, the chain from origin to registry is traceable.
2. Validate before indexing
- Content is reviewed for plausibility and authenticity before it enters the registry.
- A source with weak or absent provenance is not eligible for
governedmaturity — at mostminimal. - Sensitivity and authority are assigned from the owner’s classification, not inferred.
3. Version control
- The source is pinned to a specific version (
source_version, semver). - Consumers reference a version; they do not float to “latest” implicitly.
- The version in use is captured in the audit log on every consumption.
4. Change history
- Updates produce a new version and a change-log entry, not an in-place overwrite.
- Each change records what changed, who changed it, and why.
- A change that raises sensitivity or authority triggers re-evaluation.
5. Rollback
- A prior known-good version can be restored.
- Rollback is possible without losing the audit trail of what was in use when.
- If a poisoned or compromised version is detected, consumers can be repointed to the last trusted version.
Outcome
pass— provenance, versioning, and rollback are all in place.warn— usable at reduced maturity (minimal/operational); record the gap and the required fix.fail— provenance insufficient or rollback impossible; refuse registration / use until fixed.
Record the assessment per knowledge-audit-log-spec.
What this is not
This is operational discipline, not cryptographic supply-chain attestation. Content
signing, hash chains, and automated integrity verification are future work. The
pack_integrity_hash field in the audit-log spec is the seam where such tooling
would attach.
See also
- knowledge-source-contract — provenance and ownership requirements
- prompt-injection-in-sources-checklist
- knowledge-audit-log-spec