How it works
AI plans, reads and writes code. Code computes every number and verdict.
ClaimTrace Lab takes a research paper and tries to get the same results on an ordinary laptop. A team of AI agents does the reading, planning and coding; plain code does all the checking, so no number in the report comes from an AI's say-so.
- live has worked at least once in a real research run (test runs not counted)
- next built or planned, not working here yet
Main path
- Paper in
You give the lab a paper from arXiv.
- Reader
An AI agent reads the paper and lists each result it claims, with the exact sentence and page.
- Designer
An AI agent plans a small version of the experiment that fits on a laptop, and says in advance how close counts as a match.
- Implementer
An AI agent writes the code for that experiment.
- Laptop worker runs it
A laptop runs the code. Every number it prints is saved together with its line in the log.
- Code grades every number vs the paper
Plain code, not AI, compares each of our numbers with the paper's: reproduced, partly reproduced, or not.
- Report with log-line proof
Each claim sits next to our number, and every number links to the log line that printed it.
Side loop, alongside every step
How the team gets better from run to run.
- Controller scores agents
After each step, the controller scores the agent that did it.
- Refine, spawn or retire
Agents that keep failing are rewritten or retired; a new specialist is added when a job needs one.
- Lessons saved for the next run
What worked or went wrong is saved as a lesson.
Lessons go back to the agents at the start of the next run.
Second input
- Journal email
A reviewer emails about the paper.
- Inbox
The email shows up in the project's Inbox tab.
- Send to engine
The lab acts on the email's requests by itself.