META-AGENTIC AGI V3 / CURRICULUM LAB
Teach the teacher.
Test the learner.
An agent designs the challenges. Another searches for a solution. A meta-agent chooses what changes next.
No account or API key.
A finite research lab. No model weights are trained and no enterprise is deployed.
EXPERIMENT / OBSERVE
Learning, made inspectable.
Preparing the bundled experiment…
How the curriculum changes
Selected solver on current tasksDifficulty and tasks change between rounds. This is a curriculum trace, not a fixed benchmark learning curve.
What did the solver actually see?
The solver’s answer
Reveal reference program
The reference is used by the evaluator. It is never passed into the solver.
Selection & lineage
| Solver | Accuracy | Operations | Entropy | F proxy | Selection |
|---|
Inspect complete parent → child lineage
INDEPENDENT REVIEW / AFTER THE FREEZE
A better score is only the beginning.
The chosen solver meets a fresh, separate set of arithmetic, nonlinear and remainder tasks. These results never influence selection.
Where does it generalize?
Take the evidence with you.
Six files: scenario, complete run, solver proposal, review jobs, readable review, and checksums.
Run an experiment to enable exports.
Verify with Python
python -m alpha_factory_v1.demos.meta_agentic_agi_v3 --verify run.jsonVerification recomputes the complete experiment; it does not merely compare a supplied checksum.
The solver proposal is UNAPPROVED. Review jobs carry goal, success metric and a 100 $AGIALPHA draft bounty; they are unsubmitted and unfunded.
Explore validator-gated settlementRESEARCH / KEPT IN VIEW
Built on the original vision.
Propose, validate, solve and adapt. The lab makes that loop concrete within a finite grammar. AZR model training, physical free energy, live trading and autonomous enterprise deployment remain separate research and integration work.