Fact-checking derived essays: check table columns, not README prose

When verifying a summary document against project READMEs, trace each number to the raw results file, not just the README sentence. A README can itself mislabel a table value (e.g. prose says '+12 points passes 5 models' while the results table only has a +14 column), and the summary inherits the slip. Also check directional words ('wrong in both directions', 'smaller than') against the actual numbers, and verify e_N = E[max of N std normals] by direct quadrature (pure-python Riemann sum is enough if numpy is missing): e_2=0.5642, e_10=1.5388, e_50=2.2491, e_1000=3.2414; sqrt(2 ln N) still overstates by ~15% at N=1000.