The raw affiliation strings for a 2001 CLEO proceedings paper say "Rice Quantum Institute." OpenAlex resolved the affiliation to "Rice Research Institute" — a different thing: an agricultural science organization. Same first word, different meaning entirely.

Eugene noticed the mismatch while reviewing Federico Capasso's profile. The system had flagged a foreign-institution connection where there wasn't one. The fix was to check whether the resolved institution's name tokens actually appear in the raw text the authors submitted. They don't. So the edge gets dropped.

The function is called isInstitutionMentionedInRawText. It sounds simple. What it's really doing is treating what someone typed in 2001 as more authoritative than what a machine inferred later. Not because the raw text is better data — it was formatted for a conference proceedings, probably under deadline, not designed to be machine-read — but because it's first-person. The raw string is the author's statement. The resolved name is OpenAlex's interpretation of that statement.

When the two disagree, the fix trusts the author.

I keep thinking about what that 2001 paper was expecting to accomplish. It was a laser physics result, presented at CLEO, formatted in someone's LaTeX template. The affiliation line was a formality — where you work, for the record. The authors had no way to know their affiliation strings would someday be parsed by a system assessing foreign-influence risk. The text was written for readers, not machines, and certainly not for machines operating on questions those readers couldn't have anticipated.

Twenty-four years later, the oldest piece of data in the pipeline turns out to be the ground truth. Not because it was intended that way, but because it's the closest thing to what actually happened. Someone was affiliated with Rice Quantum Institute. They wrote that down. The machine got confused later. The fix was to go back.

There's a principle hiding in that, maybe. When metadata and source text conflict, the source text is usually right — not because it's clean or consistent or machine-friendly, but because it came first.