Unfoolable

Unfoolable is a self-improving AI agent team that can't fool itself. AI agents often say "done" when the work is wrong — our starting team did it 8 out of 9 times on hidden tasks. Unfoolable turns a one-line goal plus a messy data file into clean data: a Spec agent writes its own definition of done as executable checks, a Builder writes the code, a plain-code Watchdog catches loops, stalls and drift (continue / re-plan / stop / escalate to a human via a NEEDS_HUMAN inbox), and an independent Auditor re-derives sample rows straight from the raw file without seeing the code — if they disagree, it's not done. To improve over time, an adversarial Trickster invents new realistic data messes (and must prove each is fair with an exact inverse), a Coach proposes lessons, and a Gate keeps a lesson only if it measurably helps on fresh tasks. Results on hidden tasks never seen in training: false "done" claims dropped from 8/9 to 1/2 with the Auditor (0 false claims across 18 tasks for the base team + Auditor), self-evolution raised adversarial accuracy from 45.8% to 62.3%, and the Coach overestimated its own lessons 7 out of 7 times (avg +48 points) — the Gate rejected 4 of 7. Built from scratch today; all results, raw logs and limitations are in the repo.