# Executive summary
---
The core scaling-law claim of *When Does Adaptation Win?* reproduces. Classical LQR recovers \(G_K=A_\infty(1-e^{-\beta K})\) with \(R^2=1.000\) (\(A_\infty=0.0029\), \(\beta=0.042\)) and linear variance scaling (\(R^2\approx0.97\)–\(0.99\)). Assumption 4.3 PL curvature measures \(\mu\approx0.03\). Two-qubit CZ meta-training under 10× OOD noise shows ~54%→~97% fidelity (~+43 pp), matching the paper's >40 pp gain. Single-qubit smoke runs confirm high-\(R^2\) exponential fits and negligible gaps when \(\sigma_\tau^2\) is tiny (Corollary 4.10). Hardware: local CPU + HF Jobs `a10g-small` / prior `t4-small`; wall-clock on the order of hours; cost on the order of a few dollars.
## Scope & cost
| | This reproduction | Full replication |
|---|---|---|
| Scope | LQR + PL + reduced single-qubit MAML + CZ OOD training dynamics | Full paper schedules, denser task grids, published \(\beta\) matches on all panels |
| Hardware | Local CPU + HF `a10g-small` / `t4-small` Jobs | Multi-GPU / longer paper training |
| Compute time | ~hours (CZ GPU job + local evals) | Paper-scale multi-run suites |
| Cost | ~\$1–5 HF Jobs (est.) | Higher (full sweeps) |
| Outcome | Core claims reproduced / supported | Not fully attempted at paper schedule |
---
````html
Reproduction Poster — When Does Adaptation Win?
ICML 2026 reproduction poster — When Does Adaptation Win? Scaling Laws for Meta-Learning in Quantum Control (GATE_REPORT overall PASS).
