Not yet reviewed by Pith; the record is open.

The separation of the faint cosmological background signal from bright astrophysical foregrounds remains one of the most daunting challenges of mapping the high-redshift intergalactic medium with the redshifted 21 cm line of neutral hydrogen. Advances in mapping and modeling of diffuse and point source foregrounds have improved subtraction accuracy, but no subtraction scheme is perfect. Precisely quantifying the errors and error correlations due to missubtracted foregrounds allows for both the rigorous analysis of the 21 cm power spectrum and for the maximal isolation of the "EoR window" from foreground contamination. We present a method to infer the covariance of foreground residuals from the data itself in contrast to previous attempts at a priori modeling. We demonstrate our method by setting limits on the power spectrum using a 3 h integration from the 128-tile Murchison Widefield Array. Observing between 167 and 198 MHz, we find at 95% confidence a best limit of Delta^2(k) < 3.7 x 10^4 mK^2 at comoving scale k = 0.18 hMpc^-1 and at z = 6.8, consistent with existing limits.