ok   leafsum's shipped-arm table reproduces (84 legs, 236 windows, 8 under the 1,000-source floor, 37 merged rungs)
ok   8.17.5's published shipped-arm fit reproduces to five decimals (0.16337 / 0.01671, r2 0.99975 / 0.96421, 25 and 9 rungs, crossing 17,760.9)
ok   the other two 8.17 arms reproduce too (g64 0.15949/0.01699, noadd 0.16331/0.17482, r2 to five decimals)
ok   8.21.4's three-line fit reproduces from the banked legs: slopes 1.0595 / 0.1941 / 0.0864 and r2 0.99928 / 0.99842 / 0.99215 (the last to three decimals - 8.21 reports a width-corrected fit too), the second knee at 21,932.1 exactly and the first within 0.33%
ok   the 8.21 reduction agrees with the COMMITTED reducer harness/nttfwsum.py run on the same banked legs, both knees to 0.1 of a rung (4,230.4 and 21,932.1) - a stronger pin than the published table, and the correction to this file's earlier claim that 8.21's reducer was never committed
ok   the 8.17 rung-drop null is EXHAUSTIVE: C(37,3) = 7,770 draws, so there is no seed and no tail to be unstable in
ok   the sampled null is reproducible under its seed and moves with it
ok   --sens's claim holds: the 1,000-source floor and the 320-source one drop the same 8 windows on the shipped arm
ok   legs.json agrees with the round directories leg for leg on all four arms, so the published mirror reduces the same numbers this copy does
---
fitnull --selftest: OK - leafsum's shipped-arm table reproduces (84 legs, 236 windows, 8 under the 1,000-source floor, 37 merged rungs); 8.17.5's published shipped-arm fit reproduces to five decimals (0.16337 / 0.01671, r2 0.99975 / 0.96421, 25 and 9 rungs, crossing 17,760.9); the other two 8.17 arms reproduce too (g64 0.15949/0.01699, noadd 0.16331/0.17482, r2 to five decimals); 8.21.4's three-line fit reproduces from the banked legs: slopes 1.0595 / 0.1941 / 0.0864 and r2 0.99928 / 0.99842 / 0.99215 (the last to three decimals - 8.21 reports a width-corrected fit too), the second knee at 21,932.1 exactly and the first within 0.33%; the 8.21 reduction agrees with the COMMITTED reducer harness/nttfwsum.py run on the same banked legs, both knees to 0.1 of a rung (4,230.4 and 21,932.1) - a stronger pin than the published table, and the correction to this file's earlier claim that 8.21's reducer was never committed; the 8.17 rung-drop null is EXHAUSTIVE: C(37,3) = 7,770 draws, so there is no seed and no tail to be unstable in; the sampled null is reproducible under its seed and moves with it; --sens's claim holds: the 1,000-source floor and the 320-source one drop the same 8 windows on the shipped arm; legs.json agrees with the round directories leg for leg on all four arms, so the published mirror reduces the same numbers this copy does