All systems
Simulation · v1.1.0

Doofus

Evolve better ideas. Promote nothing that cannot survive the test.

Doofus floods the search space with competing ideas, sandboxed experiments, and quality-diversity pressure—then refuses to confuse novelty with victory. Candidates advance only when held-out benchmarks and calibrated controls survive the challenge. Evolution proposes. Evidence promotes.

Statusv1.1.0
Primary languagePython
LicenseMIT
AuthorityHuman owned

What it does.

  • Attended multi-model debate
  • Sandboxed evolutionary search
  • Held-out benchmark qualification
  • Calibrated human-owned promotion

What this system makes possible.

Doofus is more than a standalone repository. Its architecture can be adapted to a class of real operating problems.

THE PROBLEM

Innovation stalls when teams generate too few alternatives or promote attractive ideas before they survive a meaningful comparison.

WHY IT IS DIFFICULT

The system must create diverse candidates, run controlled experiments, preserve negative results, and qualify improvement on evidence it did not optimize against.

COMMERCIAL APPLICATIONS
  • Innovation and experiment pipelines
  • Strategy comparison
  • Process optimization research
  • Solution tournaments
TRANSFERABLE COMPONENTS
  • Candidate generator
  • MAP-Elites archive
  • Held-out benchmark
  • Promotion gate

Every claim comes with receipts.

Interrogate the source, seize the release, and verify the machinery for yourself.

The machine is open

Source

Skip the pitch. Open the public repository and interrogate the implementation directly.

Open repository
01Attended multi-model debate
02Sandboxed evolutionary search
03Held-out benchmark qualification
04Calibrated human-owned promotion

Need this architecture aimed at an operating problem of your own?

Design an innovation engine