Illustrative challenge figures

The calibration plot uses only the invented group counts in calibration-synthetic.json. These are not empirical NYC 311 results and expose no student, source, or hidden-test records.

Each dot summarizes 100 hypothetical requests: its x-coordinate is the group’s mean predicted probability; its y-coordinate is the number of slow_7d = 1 labels divided by 100. The highlighted bin has mean prediction 0.70 but 50/100 positive labels, so its observed fraction is 0.50. The dashed diagonal is the ideal aggregate match. A bin below that line overestimates risk in that group. Bin summaries can fluctuate with sampling; this illustration does not establish that a real model is miscalibrated or replace evaluation on held-out data.

Regenerate from the repository root with:

python3 scripts/nyc311_calibration_figure.py

The generator requires only the Python standard library and does not access the network. It preserves selectable SVG text, includes a title and long description, and uses equal physical scales for both 0–1 axes. Fixed input counts make the output reproducible. No runtime chart package, CDN, image generation, or font download is needed. SVG is used to keep axes and text crisp at all resolutions.

• calibration.svg: 720 × 562, for desktop.
• calibration-mobile.svg: 360 × 512, with larger relative labels and its callout below the plot. The course page uses this variant below 900 px to account for its sidebar, and caps the displayed narrow-layout plot at 480 px.

Suggested image alternative text: “Synthetic calibration plot: groups with mean predictions 0.10, 0.30, 0.50, 0.70, and 0.90 have observed slow fractions 0.12, 0.25, 0.40, 0.50, and 0.70. The highlighted 0.70 bin contains 50 positive labels out of 100, so its risk predictions are too high on average.”

The webpage should also provide a plain-text caption identifying the data as synthetic and explaining that calibration compares probabilities with outcomes across groups, not with an individual row’s zero/one label.

The hand-authored closure-timeline.svg (760 × 304) and closure-timeline-mobile.svg (320 × 320) show three other synthetic requests checked at day 21. Closures at day 3 and day 9 receive labels 0 and 1; no recorded closure at the mature snapshot receives label 1. Each event is positioned on a common 0–21 calendar-day scale. Days 0–7 are the outcome window; days 7–14 are the reporting buffer, not an extension of the target. These are conceptual cases, not the exact timestamps from the separate worked row.

The webpage also contains a responsive HTML diagram separating training, prediction inputs, output probabilities, and label access during scoring.