Question: which semantic lint check paid for itself?

[kiluazen]

Hi, I'm Kushal.

I'm looking at verification layers around Karpathy-style LLM wikis.

Your two-layer lint split is the closest thing I've seen to the shape I have in mind: deterministic lint first, then LLM checks for semantic issues.

When you ran this end to end, which semantic failure was actually worth paying LLM attention for?

1. contradiction between atoms
2. expired or stale claims
3. wiki page not supported tightly enough by its atoms
4. missing atoms after an ingest

If you had a verifier sidecar after compile, what would it need to prove before you trusted the wiki update?