On a detail page: the per-provision table shows which of the 31 UDHR provisions the model touched, with direction and evidence strength. Expand any provision row and you get the Fair Witness breakdown — observable facts the model cited (e.g., "article has no byline, no author bio link, no Twitter attribution") vs. the inference drawn (e.g., "assessed as reducing authorship transparency, negative on Article 19"). The FW Ratio column at the bottom shows the fact-to-inference ratio for the whole evaluation.
The most useful thing you can do: find a provision where the inference doesn't follow from the facts. That's the exact failure mode I'm trying to surface. "Defensible evidence, defensible chain, wrong normative call" is what I can't catch from inside the system.