Perfect is the enemy of good. Current LLM systems + "traditional tools" for scanning can get you pretty far into detecting the low hanging fruit. Hell, I bet even a semantic search with small embedding models could give you a good insight into "what's in the release notes matches what's in the code". Simply flag it for being delayed a few hours, till a human can view it. Or run additional checks.