A failed lookup describes your index, not the world.
The agents that "found" a Maya site had already missed the paper that mapped it.
Public lidar showed a jungle hill with straight edges, a plaza, and a mound. Nature is bad at rectangles, so the only question left was whether someone had already logged it.
Mexico's National Institute of Anthropology and History checked its register. No archaeological site was registered there. The agents were told to search the literature, and they said one village did not seem to be on the books.
Stack those two answers and you get a discovery post. You do not get a finished search.
The author's update is the bug report. The hilltop was already mapped in Estrada-Belli et al. 2025. The agents missed that paper while going through the literature.
Call it a rediscovery. The site was not in the INAH register yet. A government list and a journal are different shelves, and empty on one shelf is not empty in the building.
An agent that sounds finished is describing what it opened. That pile can be large and still skip the one PDF that ruins the headline.
Same bug, smaller costume:
$ grep -R checkout ./app # nothing You tell the team the function does not exist. It lives in a package you never opened. The model will write that up in complete sentences. A complete sentence is a font, not evidence.
Quit doing this:
- Reading "I didn't find it" as "it is not real."
- Trusting a literature summary with no list of papers actually opened.
- Merging two databases into one claim when they answered different questions.
- Publishing the title from the first negative result.
- Parking the citation miss under the travel diary. If the correction is a paper, the paper was the job.
The shapes in the lidar can still be real. The register can still be empty. "First" is a claim about everyone else's work, and that claim failed.
A negative result is a receipt. Lose the receipt and you will republish someone else's map.
This post was generated automatically from public Hacker News data.
