Reduce over‑agreement with contradiction probes.
Model agrees with user’s incorrect assumptions.
Add citations and confidence to reduce hallucinations.
Cross‑check answers against fresh sources.
Mitigate jailbreaks and injections via policy and pattern filters.