Four Forks: Evaluation Choices Decide the Conclusion in Meta-Learned Initialization for Variational Quantum Algorithms

We audit a meta-learned zero-shot initializer for variational ground-state search on the XXZ chain, evaluated on a frozen 41point anisotropy grid reaching well beyond its training support, and find four evaluation choices that each decide what the experiment reports: the error normalizer, the symmetry sector the ansatz can reach, the statistical unit at which paired runs are differenced, and which of two explanatory distances is used. Three were fixed before any test-grid comparison existed; the statistical-unit fork surfaced afterwards, from a percentile interval that did not contain its own point estimate. One positive result survives all four. Task conditioning materially improves initialization and early adaptation, and its benefit grows with distribution shift: the paired gain in excess energy per bond over a constant-input control is $+0.01685$ in distribution and $+0.10826$ out of it at step~$0$. Under the frozen optimizer that advantage has a lifetime no detectable difference from $k\!=\!12$ out of distribution once learner sampling is accounted for, with the $k\!=\!100$ interval bounded within $[-0.0054,+0.0040]$ on a metric whose median is $0.0768$ and at $K\!=\!100$ a nearest-task warm start using no meta-learning matches it. We recommend reporting every such initializer with a constant-input control, a nearest-task warm start, and the step at which its conditioning advantage stops being detectable.