Running tally — classifications
What you just did has a name
The boundary between drift and selection is |s| ≈ 1/Ne.
Below that, selection is too weak to overcome the sampling noise of
reproduction, and the allele behaves as if neutral. Above it, selection
dominates. In between, you see a rising (or falling) trend with
substantial scatter — the combined regime.
Round 1 was pure drift in a tiny population: the trajectories fan out with no directional trend. Round 2 was selection in an essentially infinite population: the trajectories are a single deterministic line. Round 3 was the combined regime — the trend is visible but any single replicate could mislead you. Round 4 used Lenski's LTEE Ara+1 fitness record: relative fitness is not an allele frequency, but the statistical question ("is the trend above what random reshuffling would produce?") is identical. Round 5 used a real Florida Scrub-Jay SNP trajectory across 23 years — the observed Δp is consistent with drift at the estimated Ne.
Selection is the correlation; drift is the scatter. Selection shows as a consistent covariance between allele and fitness across many individuals and generations. Drift is the variance around that trend. In small populations or over short timescales, the variance dwarfs the trend — so drift is always the null, and selection must be demonstrated.