Problem
VQE benchmarks are usually iteration-bound. That hides the real cost: gradient evaluation, transpilation, optimizer overhead, and hardware-noise stochasticity.
Contribution
A single-file research harness that fixes wall-clock budget per problem size, hands the experimenter one editable train.py, and logs every run as a row in results.tsv.
$$ \min_{\boldsymbol{\theta}} \; \langle \psi(\boldsymbol{\theta}) | \hat{H} | \psi(\boldsymbol{\theta}) \rangle \quad \text{s.t.} \quad t_{\text{run}} \le 2^{n-2} \text{ s} $$
Eq. 1 — VQE under a fixed wall-clock budget.
Status
Feature-complete on synthetic Hamiltonians; current work is on noise-aware extensions and agent-driven search loops.