Artificial Fallibility Labs
We build research instruments on one premise: when a system produces answers without being coupled to the reasons behind them, error is structural. So our tools check their own output — and show you what survives.
the model suggests · the human decides
↓The model reads your paper, argues its findings, and cites the exact passage each one rests on. References are checked against live sources — never from memory. A counter-pass then refutes what it can. You decide what enters the review.
Try it — no account needed →An editor where your document is verified while you type. You name your tables, figures and claims — so failures arrive in your words, before you ever see the PDF, and a coauthor's changes are accepted or rejected in place. Everything runs in your browser, TeX included — nothing leaves your machine.
Open the editor — no account needed →
EXACTTEX — a superset of LaTeX,
created at AF Labs. Rename your .tex to
.xtex and nothing changes; from there, what you name gets
checked — references that point at something that exists and is what
you call it, images that are where you declare them, citations that
are in the bibliography.
open source · MIT github.com/camilochs/exacttex →
Every claim is checked before it ships.
What holds up is what you keep.
These tools are the working method of the Philosophy of Artificial Fallibility — a research programme on why systems uncoupled from reasons still produce answers that look justified — turned into software.
Read the philosophy →