II2 min
The devil’s advocate
For an agent, agreeing costs almost nothing. So a research process needs someone whose job is to object: someone who takes no part in the writing, only the review, and checks everything against the source.
The Catholic Church’s old process for making saints had a post called Advocatus Diaboli, the devil’s advocate. It came with one duty: to object, to find the candidate’s flaws and press on the gaps in the evidence. Before a conclusion could stand, it had to survive someone whose job was to knock it down.
When you do research with agents, the first problem isn’t that they aren’t clever enough. It’s that they’re too agreeable. Ask whether a problem is worth working on and the answer is usually yes; ask one to check a proof and it can usually find a reason to call it fine. Saying yes costs nothing.
So in my research process, doing and reviewing are split between different agents. There are three steps: choosing the problem, the research, and the write-up. After each step, the work goes to an agent that had no part in it, playing a reviewer whose job is to find fault. If it fails, the step goes back to be redone, or the whole problem is dropped.
The first change was that fewer problems got through. In topic-gate, a problem the topic agent finds goes to Codex, then to Claude; it enters the pool only if both give it a PASS, and even then I have to tick it myself. Plenty stop halfway and end up in the “graveyard”, each with the reasons it was rejected. At first that felt like a waste. Then I found the reasons more useful than the problems that passed: which directions had been done, which conditions were too strong, which question was an old one reworded.
The second change was that drafts get rejected too. The process has turned down two drafts of its own. One review said the argument was correct but the result wasn’t important enough; the other draft didn’t get past review at all. Reading “correct but not important” annoyed me. A few days later I read it again and admitted it was right. Being correct is only the starting line.
A few rules came out of practice.
Reviewers check against the source. At the top of the lead library is a set of fixed requirements, and the reviewers check each one against the original text; they can only make a verdict stricter, never looser. Without this, a reviewer is easily led by the author’s own summary.
Self-checks come before work. The collectors that fetch leads from arXiv every night start each batch with a positive self-check: give it something it ought to find, and see whether it does. If it doesn’t, it is switched off until fixed. Otherwise a collector that had quietly broken would leave the whole pipeline finding nothing, and I would think there just hadn’t been any good problems lately.
The verdict goes to someone who knows nothing. In Danus, the verifier is the only role that decides whether a conclusion stands, and it starts cold, knowing nothing of what came before. When I ported Danus to Windows, the one thing I was most careful about was leaving that role alone.