Check the numbers in old research papers with two agents
Economic historian Johan Fourie gave Claude and Codex his hard drive and asked them to check 30 of his own papers from 2010 to 2025. They extracted 11,005 numbers and agreed on 102 headline claims. Of those, 52 held intact, 17 held with a different size, 12 weakened, 5 fell and 16 could not be tested. It took about 29 hours and $269, but code survived for only 18 papers, and one question of interpretation stayed with him.
Done with CodexModerate
What was done
Economic historian Johan Fourie gave Claude and Codex his hard drive and asked them to check 30 of his own papers from 2010 to 2025. They extracted 11,005 numbers and agreed on 102 headline claims. Of those, 52 held intact, 17 held with a different size, 12 weakened, 5 fell and 16 could not be tested. It took about 29 hours and $269, but code survived for only 18 papers, and one question of interpretation stayed with him.
How to do it
- Paste the prompt into the tool, and replace the text in brackets with your own.
- Read the answer through. Do not send, delete or publish anything until you have seen it yourself.
- Check the original before you trust a figure, a name or a date.
Try it with Codex
Here is the folder with my papers, data and code. First extract every reported number, and make a locked list of the headline claims before you calculate anything. Rerun the code where it exists, and rebuild the analysis from the data where it is missing. Show me every claim that does not hold, with file and table, before you change anything. You are done when every claim is marked as holds, weakened, overturned or not testable.
Johan Fourie on Substack ·