Tutorials

The notebooks under notebooks/, executed top to bottom when this site was built: every figure, table and printed number on these pages came out of that run, not out of the outputs saved in the repository.

notebook	what it shows
0 · Start here: a decision from logs	the whole loop on one everyday problem: the trap in the logs, the effect, the plan, and how far to trust it
1 · Causal vs predictive control	the headline: under confounding, predictive control drives the state the wrong way (to −20, target +2); causal control matches the oracle
2 · Learning what physics misses	hybrid dynamics + system ID: recover an omitted cubic term; multi-step training cuts drift from noisy measurements
3 · The causal inference toolkit	adjustment · IV/2SLS · Double ML · sensitivity · refutation, side by side
4 · Scientific control & offline safety	flatten an epidemic curve down to a capacity cap; pessimism vs a greedy controller
5 · The benchmark scoreboard	regret vs oracle across three tasks — CHC matches the oracle under confounding and beats the greedy baseline on support-shift
5 · Robust control under hidden confounding	when no adjustment set exists: a sensitivity level Γ → pessimism radius → minimax action, and what it costs when there is no confounding or the bias runs the other way
6 · Adaptive cruise control from fleet logs	relatable end-to-end: confounded fleet logs (Simpson's paradox → adjustment or IV → control)
7 · The LaLonde validation	real data, experimental ground truth: the naive estimate flips sign; adjustment and Double ML land inside the experiment's 95% interval
8 · Splitting an advertising budget	three channels, one budget: why the best split equalises the next euro's return, and when TV switches on
9 · Heating a room on a winter night	a thermostat's logs, the weather that confounds them, and a night plan that keeps the room warm

Run them yourself

In a checkout of the repository:

uv sync --group notebooks
uv run --group notebooks jupyter lab

Each .ipynb has a paired jupytext .py source beside it. Notebook 7 downloads the Dehejia–Wahba files from NBER on first use into notebooks/data/, which git ignores; nothing is committed. To rebuild these pages, just docs executes them all and then builds the site.