Contention resolution schemes (CRS) transform random sets that are ex‑ante feasible for a constraint into sets that are ex‑post feasible. They are central to Bayesian and stochastic optimization such as prophet inequalities and stochastic probing. Full knowledge of the underlying distributions is often unrealistic, motivating the study of sample‑based CRS. This work investigates $k$‑uniform matroids under both uniformly random and adversarial arrival orders, comparing the balance guarantees of sample‑based schemes to the optimal guarantees achievable with full distributional knowledge.\
For the single‑item case ($k=1$), one sample per element suffices to obtain an optimal $(1-1/e)$‑balanced ROCRS. In contrast, under adversarial arrivals—even when the order is known—a single‑sample OCRS cannot be $1/2$‑balanced, separating sample‑based contention resolution from prophet inequalities in the single‑sample regime. Moreover, two samples are both necessary and sufficient to achieve the optimal $1/2$ guarantee, and a single‑sample $0.48$‑balanced scheme is also presented.\
For general $k$, a simple single‑sample ROCRS attains the optimal balance ratio $1-e^{-k}k^k/k!$, matching the full‑distribution bound. Under adversarial order, a single‑sample scheme achieves a $(1-O(1/\sqrt{k}))$‑balanced guarantee, which is asymptotically optimal in $k$.\
Finally, the paper characterizes the minimum number of samples required to match the optimal balance. With random arrivals, one sample suffices for $k=1,2,3,4$, while zero samples are enough for $k\ge5$. For adversarial arrivals (even when known in advance) and any $k\ge2$, no constant number of samples can achieve the optimal balance.\
Review: The paper provides a thorough analysis of how sample complexity governs the performance of contention resolution schemes, delivering both optimal constructions and tight lower bounds. These results give clear guidance for designing stochastic optimization algorithms when only limited sampling is feasible.