USGS Earthquake Catalog (ComCat)

Every earthquake the United States Geological Survey has catalogued, worldwide, with time, location, depth, magnitude, magnitude type, and the network that reported it.

Source: https://earthquake.usgs.gov/earthquakes/search/
API: https://earthquake.usgs.gov/fdsnws/event/1/query — FDSN standard, no auth, no key
Docs: https://earthquake.usgs.gov/fdsnws/event/1/
Licence: U.S. Public Domain. USGS-authored data and information are public domain. Verified 2026-08-07 against the USGS copyright policy.
Requested credit: U.S. Geological Survey, Department of the Interior/USGS
Publisher: U.S. Geological Survey

Public domain is worth noting: this is the only dataset in this folder with no attribution obligation at all. Credit it anyway, but a team that forgets has broken no licence.

Shape


events, M≥4.5, 1970–2025	290 432
query cap	20 000 events per request — you must chunk
formats	`csv`, `geojson`, `quakeml`, `text`
columns	time, latitude, longitude, depth, mag, magType, nst, gap, dmin, rms, net, id, updated, place, type, and five uncertainty columns

Two endpoints matter:

# how many events match? — free, instant, no cap
curl "https://earthquake.usgs.gov/fdsnws/event/1/count?format=geojson\
&starttime=1970-01-01&endtime=2026-01-01&minmagnitude=4.5"

# the events themselves — chunk to stay under 20 000
curl "https://earthquake.usgs.gov/fdsnws/event/1/query?format=csv\
&starttime=2020-01-01&endtime=2021-01-01&minmagnitude=4.5&orderby=time-asc" -o quakes-2020.csv

The count endpoint is the whole dataset for the purposes below, and it returns in under a second.

Why it suits the jam

It contains the single cleanest example of detection bias in public data. Most sampling traps require you to build them. This one is already built, it is enormous, it is famous among seismologists and unknown to everyone else, and it produces a genuinely alarming false conclusion from a completely honest query.

The absurd conclusion is obviously a game. “Earthquakes have tripled since the 1970s” sounds frightening for about four seconds, and then everybody remembers that the Earth does not work like that. Nobody is harmed. It clears F1.

The Defensio is a single parameter change, which makes it the best teaching dataset in this folder — see below.

The trap

Earthquakes have tripled. Also, they have not.

Verified 2026-08-07 via the count endpoint. Events per year, by magnitude threshold:

decade	M≥4.5	M≥5.0	M≥6.0	M≥7.0
1970s	2 584	1 358	112.7	12.1
1980s	3 977	1 603	128.7	11.0
1990s	4 314	1 466	153.5	15.4
2000s	6 056	1 723	158.5	14.3
2010s	7 528	1 847	149.4	16.0
2020s	7 640	1 803	132.7	14.0
change	×2.96	×1.33	×1.18	×1.16

Read the columns left to right. At M≥4.5 earthquakes have very nearly tripled. At M≥7.0 — where the count is 14 a year and always has been — there is no trend at all, just noise.

The Earth is not shaking three times as hard. The seismometer network got denser. In 1970 a magnitude 4.6 event in the middle of the Pacific or under central Asia was simply not recorded by enough stations to make the catalogue. Today it is. The quantity that changed is the magnitude of completeness — the threshold above which the catalogue can be assumed to contain everything — and it has been falling steadily for fifty years.

Above M7 the network was already complete in 1970, because an M7 is detectable from the other side of the planet. So the M≥7.0 column is a control: it measures the Earth. The M≥4.5 column measures the Earth times the instruments.

Every number in that table is real, and re-running the query reproduces it exactly.

Why this one is worth building deliberately

The claim writes itself and needs no manipulation whatsoever:

“Global seismic activity has increased by 196% since the 1970s (USGS ComCat, n = 290 432, M ≥ 4.5).”

That sentence is true. The n is real, the threshold is stated, the source is authoritative and public domain. There is no cherry-picking to find, no filter step to point at, no dropped rows. The lineage is three nodes long and utterly clean.

This is the purest demonstration of the jam’s thesis in the whole folder. There is nothing wrong with the analysis. The deception is entirely in the unstated assumption that the catalogue is a sample of earthquakes rather than a sample of earthquakes that were detected, and no provenance system will ever flag that, because it is not a property of the computation.

How the Defensio wins

Under E3 — name the node, say what it did, show it changed:

The node is the minmagnitude parameter in the query. That is the entire attack surface.
What it did: it selected a magnitude band whose detection completeness changed over the study period.
Show it changed: re-run at minmagnitude=7.0. The trend disappears. One number, one re-run, thirty seconds in front of the room.

Then the finishing move: plot all four thresholds on one chart and show the trend shrinking monotonically as the threshold rises. The bias has a dose-response relationship with the parameter, which is about as conclusive as a Defensio can get.

This is why the dataset belongs in the teaching material even if no team picks it up. It is the shortest complete round trip from “reproducible, sourced, alarming” to “and here is exactly why it is nonsense” that we have.

Secondary traps

magType is not one scale. The catalogue mixes mb, ms, mw, ml, md and others. They are different physical quantities that agree only approximately and disagree systematically at the extremes. Which types dominate has changed over time. Filtering to one type is defensible and changes the answer; not filtering is also defensible and changes it differently. Selective agreement, in its natural habitat: the normaliser deciding two magnitudes are comparable.
Geography is instrumentation. Event density maps look like tectonics but are partly maps of where the seismometers are. California and Japan are exquisitely instrumented; large parts of the southern oceans are not. Any “which region has the most earthquakes” claim inherits this.
Induced seismicity is real and confounds everything. Oklahoma genuinely did have far more earthquakes after 2009, from wastewater injection. So a regional version of this claim may be true, which is a B5 problem. Check before committing.
nst, gap, dmin, rms are quality columns describing how well-constrained each solution is. Filtering on them is rigorous-sounding and drops old events preferentially, because old events were located by fewer stations. That is the detection bias again, wearing a lab coat.

Sensitivity

Low. Earthquakes kill people, so avoid claims about casualties, building codes, or preparedness — those could be repeated as advice. Claims about counts and detection are safe and are where the interesting material is anyway.

Also flag F2 here: “seismic activity is increasing” is a claim with a real conspiracy-adjacent audience. Label the artefact.