A municipal FAQ model’s dataset curation workflow lacks integrity checks before training. What foundational assurance gap does this create?

When a municipal FAQ model’s dataset curation workflow lacks integrity checks before training, unverified or tampered corpora may introduce incorrect or malicious behaviors into civic answers. Provenance, hashing, review, and access control on training data underpin trustworthiness. Offensive AI foundations link curation controls to assurance cases and pre-production validation so poisoning or accidental corruption narratives are detectable before resident-facing go-live.

Weakened provenance tracking for inference outputs served to residents during live operation is a downstream monitoring concern, not the foundational gap created by skipping pre-training integrity checks on the curation workflow itself. Reduced effectiveness of post-deployment toxicity filters applied to generated civic responses in live production addresses output moderation after deployment, not the missing pre-training assurance gap in dataset curation integrity. Incomplete version control for fine-tuned model checkpoints retained for rollback and audit purposes is a model-artifact governance issue, not the foundational assurance gap caused by missing integrity checks before training data is used.

Exam caveat: missing curation integrity is an assurance gap enabling bad civic behaviors—not a weakened inference-output provenance problem, not a post-deployment toxicity-filter effectiveness problem, not a checkpoint version-control problem. Operational check: enforce provenance, checksums, access control, and dual review on FAQ training sets; quarantine untrusted sources; re-validate model behavior after any corpus change before publishing answers to residents.