A model predicts fluent words, not true ones. This sieve strips a passage to its proper-noun phrases and dates and runs a live Wikipedia audit, asking one thing of each anchor: does it actually exist?
Live preview · launch for the interactive version
Highlight the hard factual anchors and dim the fluency that hides them.
Each anchor is queried against Wikipedia — invented names fail here.
A passage can pass existence and still be false — that's the next sieve's job.
Predict which anchors in the passage are real.
Replace a real name with an invented one; re-run.
Which anchors resolved, which came back unfound?
Say what failed: a fabricated entity, or a real but unverifiable one.