Rung 4 — What Isn't There

Five dives, built and ready to run. Data in data/, rebuildable with python3 fetch.py.

The last rung, and the dive changes shape. A senior does not get a chart. They get a claim somebody built on one, and the question is whether the evidence carries it.

That is the professional turn made routine: six weeks before the capstone asks them to recommend a site to a stakeholder who "doesn't speak GIS", they spend twenty warm-ups finding out what a recommendation is worth.

Q1 the trap. Where does this series start and end, and who chose that? Q2 the finding. Does the evidence reach the claim? Q3 the turn. What would you need before you would sign it?

Every claim below is one a reasonable person would make from the chart. Every one fails.

---

4.1 · "These counties have clean air."

absence_ca_counties_air_monitoring.csv — 58 California counties.

Draw the choropleth: median AQI by county. Six counties come out blank.

Q2. What does blank mean?

A viewer reads white as good. It means nobody measured.

261,239 Californians live in a county with no air-quality monitor, across 11,564 square miles. Napa has more people than seven monitored counties and no monitor at all.

Q3. You are recommending where to site a facility that affects air quality. Two candidate counties: one with a middling AQI, one blank. Which is safer to build in, and how would you find out?

The blank county is not the safe one. It is the unmeasured one, and a recommendation that treats those as the same thing has already failed.

---

4.2 · "The cluster is here."

hexbin_la_311_2km.csv · hexbin_la_311_5km.csv — the same 1,500 LA 311 requests, binned twice.

Q1. Not one data point changed. The bin size changed. The cluster moved seven kilometres and across the city. Q2. Which map is right? Q3. You are about to tell a client "demand concentrates here." What do you have to publish alongside that sentence for it to mean anything?

This is the Modifiable Areal Unit Problem, and it is the library's own named trap for this shape. Nobody chose those hexagons for the question — somebody chose them for the picture.

A note on tooling, because it is the point. Neither Flourish nor CODAP has a hex-bin template, so fetch.py computes the bins itself. That was necessary rather than incidental: a dive about what the bin size does needs the bin size to be ours to change. Re-run it at 1 km or 10 km and watch the answer move again.

---

4.3 · "Every region is warming."

multiples_gistemp_zonal.csv — NASA GISTEMP by latitude band, 1975–2025.

The global line rises. The claim looks safe. Draw the eight bands as small multiples.

Q2. The Arctic is warming forty-six times faster than the Southern Ocean band. On the global average, both are one line. Q3. The claim is true of the planet and false of a band containing an ocean. Is it a true claim?

The eye averages panels, and an average is built to hide exactly this. A student who argues the claim is defensible is not wrong — that is the discussion.

Worth knowing, not required: the Southern Ocean overturns, bringing deep cold water to the surface. The flat panel is physics, not error, which is why the honest answer to Q3 is "yes, and say which band you mean."

---

4.4 · "Southern California earthquakes are declining."

noise_socal_m4_by_year.csv — M4+ per year, 1995–2024.

`` 95:7 96:7 97:12 98:8 99:67 00:6 01:8 02:7 03:11 04:5 05:14 06:5 07:9 08:8 09:9 10:11 11:3 12:7 13:6 14:8 15:5 16:6 17:0 18:4 19:3 20:6 21:2 22:2 23:4 24:15 ``

Stand in 2022. The last four years read 6, 2, 2 after a decade averaging nine, and 2017 recorded none at all. The decline looks real.

Q2. Then 2024 returns 15. Q1. The mean is 9.1. A first-year noise model puts the band at roughly 3 to 15 — so 2021 and 2022 sit outside it and the decline looks significant.

That model is wrong, and finding out why is the dive. Earthquakes are not independent events. They arrive in families: 1999's 67 is one magnitude 7.1 and its aftershocks, not sixty-seven separate facts. When events cluster, the real spread is far wider than the simple arithmetic predicts, and a band computed the easy way calls ordinary years extraordinary.

Q3. You have thirty numbers and a claim. What would convince you the decline is real — and can thirty numbers ever do it?

A zero is not a gap. 2017 is written as 0 in the file, not omitted, because a missing row reads as nobody looked and a zero reads as we looked and found none. Those are different claims and the file has to say which.

---

4.5 · "Earth-like planets are rare."

void_exoplanet_mass_period.csv — 5,981 confirmed planets with mass and orbital period.

Plot mass against orbital period. There is a large, obvious, empty region, and Earth is in it.

Q2. Earth is 1 Earth mass at 365 days. The archive contains nothing like Earth. Does that mean planets like Earth are rare?

No. It means we cannot see them. A transit search needs repeated transits, so a long-period planet gives a mission two or three chances in its whole lifetime. A radial velocity signal shrinks with mass and with period. The empty corner is the shape of our instruments, printed onto the sky.

Q3. Somebody uses this chart to argue Earth is unusual — perhaps that we are alone. What is the one sentence you would add to the chart to stop that reading?

Absence of evidence, drawn. Four years of "absence is data too" arrive here as the largest question a person can ask, and the answer is that the hole in the data is a picture of the telescope.

---

The MAKE dive

You are handed a claim and the chart behind it. Produce the honest version.

Not a rebuttal. The chart the claimant should have drawn — same data, same finding if the finding survives, with whatever the original left out put back on the page.

Then one paragraph: what you changed, and what your version can still not tell anybody.

That second half is the capstone's whole standard arriving early. A recommendation that names its own limits is the one a stakeholder can act on.

---

What changed on contact with the data

Two dives in this rung are not the ones the bank specified, and both changed because the data refused:

• 4.5 was specced as a bivariate choropleth of trade-area cuisine — "no competitor serves this cuisine here." The LA 311 version of that test found zero empty cells: all 15 council districts carry all 12 request types, so there was no absence to teach from. The exoplanet void carries the identical lesson and is not manufactured.
• 4.1 was specced against USDA FARA low-income low-access tracts. Both published file paths return 404. The California monitoring gap needed no new source and is arguably sharper, because the absence is of the instrument rather than of a shop.

Provenance

SOURCES.md. This rung adds NASA GISTEMP zonal means and the hex-binning computed in fetch.py; every other publisher was already banked.