ShotPilot: when is adaptive measurement worth its cost?
Simulation study for the Quantinuum Singapore Grand Challenge. This notebook
evaluates uniform, variance-only, cost-aware and guarded allocation against an
oracle reference. All pilot shots and per-job overheads count against the same
total budget. Production samples are independent of the pilot, avoiding bias
from reusing observations that determined the allocation.
The guarded controller skips pilot acquisition when its predicted cost exceeds
20% of the budget. Otherwise it uses regularized pilot variances and falls back
to equal allocation if the predicted variance gain is under 10%. These settings
are declared in advance; this notebook does not tune them on evaluation results.
A guardrail is a hypothesis, not a guarantee of outperforming a baseline.
Five prepared H2 states, two synthetic noise levels, three cost regimes and three
budgets form a 90-case suite. All cases, including losses, are displayed.
