What it is
QuantumMytheme is an open-source bench for verifiable quantum design. You point a capable model at a hard problem — prepare a quantum state, minimize a Hamiltonian, design a chip topology, build a feature map — and a hermetic, numpy-only judge re-simulates the answer and ACCEPTs or REJECTs it. No human in the scoring loop, no model grading its own homework, and every verdict reproduces on a laptop with no quantum hardware. (I developed the whole thing on a Raspberry Pi.)
How it works
A submission is a self-contained proof bundle — the circuit, the constraints it was solved under, and the number it claims. The judge re-derives everything from scratch through four gates, first failure wins:
- Structure — the circuit respects its qubit count, depth budget, native gates, and coupling map
- Reproducibility — re-simulating reproduces the claimed number; a bundle can claim fidelity 1.0, but the judge recomputes it and rejects the lie
- Performance — the recomputed result clears the bar and beats the classical baseline
- Anti-overfit — it also passes a held-out check the model never saw: a hidden observable, workload, or test set
The answer key lives only with the judge, so a model has to genuinely build the thing, not parrot a target. There’s also a 39-slice curriculum that climbs from a single bit and a single qubit up to thirteen re-runnable landmark experiments — Bell’s inequality, teleportation, error correction, the RSA→Shor bridge — each computing its headline number live in your browser.
The honest pivot — and why it matters
I started this convinced quantum would help make AI more efficient. Working through the research — paired with Claude (Opus 4.8) running adversarial, multi-agent workflows — I got talked out of my own premise. The honest finding, which the site now states plainly: a quantum computer will not make today’s LLMs faster, cheaper, or greener. The near-term efficiency gains are classical.
So I reframed the whole thing. QuantumMytheme isn’t a quantum-hype site — it’s a verifiable-efficiency referee: one re-checkable yardstick for where machine intelligence actually gets more efficient, with quantum design as the hardest case to verify, which is why I started there.
That’s the same thread running through everything I build — the Robot Registry Foundation, and the rest: claims that decide things should be legible, signed, and re-checkable by a stranger — not taken on faith. This one points that conviction at the most hype-prone field I could find.