I ran the pipeline on 15 Stockholm startups, read every failure, and changed it twice. Each version was scored on the same startups.
Results
What each metric means
Backed by source. Of the facts whose cited page could be read, the share the page fully states. Judged by a second model call that only sees the page text.
Not in source. Facts the cited page doesn't state or contradicts. This is the number that matters most: a confident fact with a source that doesn't say it.
Unreadable source. The cited page couldn't be read without a browser (blocked, or rendered by JavaScript), so the fact couldn't be checked.
Own-site sources. Share of facts cited to the startup's own website rather than press or directories.
One-liner correct. The brief's summary of the business compared with a one-line reference I wrote from each startup's site, judged by a model call.
Limits
15 startups is small, and each version ran once: a difference of one or two facts moves a rate a lot. The checker is a model too, so its verdicts can be wrong. The reference one-liners are mine. Treat these as signals for where to look, not as a benchmark.