The endpoint sweep of Section 4 as a word RAM approximation scheme. Read the instance and the accuracy $e$ (in the place of the threshold), sort the jobs by start time, build the sorted arrays and the due-date order (as in the exact program), zero the weight of every unfit job, take $k = max 1 (w_max' / (e n))$, replace every weight $w ≥ 1$ by $(w - 1)/k + 1$, set the threshold $W = 2 e n²$, run the exact endpoint sweep ($W3Sweep.core$, unchanged) on the rescaled weights, scan the finished row $TB[0 … W]$ for the last finite cell $W'$, and write $k (W' - n)$ if $k > 1$ and $W'$ otherwise ($F5Math.fptasOut$). The cost is that of the exact program with threshold $2 e n²$, that is $O((n+1)² (e+1) · 2^ω · (n+1))$, plus the sort; the guarantee is $F5Math.fptas_value$.