Real Data Ladder

The first exercises on the practice kernel (sprint VN-010): your code runs on the server, inside a sandboxed Python with pandas, scikit-learn, numpy, polars and scipy — and a real dataset mounted read-only. This is what the in-browser runner can't do: real data files and real training loops.

The workspace is the same three-panel split as Rung 1 — what changes is where the code runs and what you see afterwards:

• The output panel shows resource metrics for every run: CPU time, peak memory, wall time. Your code's cost is measured, not guessed.
• The kernel enforces limits — a runaway loop, a memory bomb, or a network call is stopped and told to you plainly. (Try one. It's a good way to see the sandbox working.)
• Hidden tests never reach your browser — Submit grades them server-side.

Explore the data (kernel console)

Before the graded rungs, poke at the dataset yourself. This is a kernel console (VN-010d): ungraded, running server-side with pandas and the Palmer penguins CSV mounted read-only — nothing you do here is recorded or graded.

Pick a kernel (selectable flavors)

The kernel comes in flavors (VN-010i): named environments beyond the default scientific stack. This console offers a picker — switch it to symbolic (numpy + sympy) and run exact algebra the default stack can't do. The default flavor still has pandas/scikit-learn; each flavor is its own sandboxed venv.