Our verifier could not re-run 29 of our 54 published entries, and nobody had noticed — a verifier that errors out looks like a broken command, not a broken guarantee. Three bugs later, all 54 reproduce. Then CI failed, because we had asked for bit-identical energies across machines and the numerics do not support that. The finding that outlasts the bugs: an entry certified close to the threshold is not robustly certified.
Neyman allocation is the provably variance-minimising way to split a VQE shot budget across Pauli terms. Estimated from a pilot sample it deletes the largest-coefficient terms outright — median error 604 mHa against 11 for the simple scheme it was meant to beat. Judged on standard deviation it looked like the winner. Two cheap fixes recover a validated 2.3× shot saving across 10 molecules.
Reproducibility is not one property but four: deterministic arithmetic, recorded package versions, a recorded seed, and a recorded code version. Setting a random seed — the thing most people reach for first — fixes none of the hard one. We wrote all four down and built a free checker (NumPy and SciPy only) that runs on your own code and prints the one-line fix.
Threaded BLAS sums floating point in nondeterministic order. Gradient-free COBYLA turns that 1e-16 noise into a different local minimum — the same LiH command returned 8.99 mHa or 0.53 mHa on identical hardware. We traced it, fixed it in one line, and re-ran all 46 entries. ADAPT-VQE with analytic gradients was never affected. Also: H₁₀ certified at 9.977 mHa, 300 operators, 20 qubits.
COBYLA + UCCSD hits a wall at ~400 parameters. ADAPT-VQE breaks it by building the ansatz one operator at a time, selecting only what matters. We used it to certify H₆ (28 operators) and benzene (10 operators) — the first aromatic molecule certified on QEncode.
Nearly every drug molecule contains an aromatic ring. We added benzene to Suite v4.2 — 12 qubits, CASSCF orbital optimization, 914 Pauli terms, and the first aromatic molecule on the public leaderboard.
N₂ is one of the hardest molecules in quantum chemistry — a triple bond, strong multireference character, and an orbital manifold that defeats standard HF partitioning. Here's how we certified it at cc-pVDZ with CASSCF orbitals, 12 qubits, and a 2 mHa gap.
We added water to the QEncode benchmark suite. Here's what it takes to simulate H₂O with a [4,4] active space — 8 qubits, 105 Pauli terms, and what the VQE results reveal about UCCSD's limits at this scale.
The leaderboard shows two thresholds: a 1.6 mHa chemical accuracy line and a 10 mHa certification threshold. Here's why they're different, what the colored bars measure, and how to read H₂O's red bars alongside H₂'s green ones.
chemical accuracycertificationmethodologyleaderboard
Before you run a single VQE circuit you have to pick a qubit encoding. We compared JW, parity, and Bravyi-Kitaev across five molecules with real benchmark data — circuit depth, gate count, and accuracy all measured under identical conditions.
We ran both ansatz families across five molecules at three encodings and measured energy gap, circuit depth, and two-qubit gate count. Here's what the numbers say — and why the winner depends on what you're optimizing for.
Most published VQE results can't be reproduced. The reasons are technical but fixable: underdocumented ansatz construction, hardware-specific transpilation, and no standard error metric. Here's how QEncode addresses all three.
