Here are the three questions I think this dataset answers best. I used the per-species summary (`Iris_by_Species`, 50 specimens per species) to check that each one can be answered. ## 1. Which measurements best separate the three species: petal or sepal? The petal measurements separate the species much better than the sepal ones. - **Petal length:** the average is 1.46 cm for setosa, 4.26 for versicolor and 5.55 for virginica. Setosa's range (1.0–1.9) does not overlap with either of the others. - **Sepal width:** the averages (3.42, 2.77, 2.97) are close together, and the ranges overlap heavily (setosa 2.3–4.4, versicolor 2.0–3.4, virginica 2.2–3.8). - **Sepal length:** versicolor (4.9–7.0) and virginica (4.9–7.9) overlap almost entirely. A follow-up query could rank each feature by between-species variance divided by within-species variance. ## 2. Is a simple threshold on petal size enough, or do you need a multivariate model? - **Setosa:** a single rule such as petal length below about 2 cm (setosa's maximum is 1.9, versicolor's minimum is 3.0) isolates it perfectly. - **Versicolor vs. virginica:** a single threshold is not enough. Their petal lengths overlap at 4.5–5.1 cm, and their petal widths overlap at 1.4–1.8 cm. To measure how much a one-feature rule misclassifies, I would count the specimens that fall in those overlap bands. I would then test whether combining petal length and width closes the gap. ## 3. How variable is each species, and are there outliers or borderline specimens? - **Spread:** virginica has the widest sepal length range (4.9–7.9). Setosa has the widest sepal width range relative to its mean (2.3–4.4). - **Borderline specimens:** a few versicolor and virginica flowers sit in the overlapping petal ranges. The row-level table (`raw.Iris`, key `Id`) would let me list them and test whether they look like outliers or mislabels. - **Why it matters:** this shows which specimens a classifier will get wrong. It also helps when choosing between simple rules, clustering and supervised models on a small dataset where one class separates cleanly and the other two do not. Each of these can be checked with SQL on `raw.Iris`, such as overlap counts, z-scores within species and correlations between measurements. I can run any of them next.