Skip to content
Science Journaling Club

INTERACTIVE MODELFIELD NOTEPALEOBIOLOGY

The Fossil Detection Bench

A live model accompanying “The Dinosaurs Were Fine. The Rocks Weren't.”

Read the full field note

Study
Field note
Review
Peer-edited by the club review board
Article
13 sectionsabout 24 min
Files

The whole argument in the field note rests on one relationship: what you observe is true occupancy multiplied by detection probability, not true occupancy alone. Model 1 lets you feel that out on a grid of sites you can fully see the truth of. Model 2 puts the same mechanism on a Cretaceous timeline and asks whether a flat, healthy dinosaur record could still produce the famous decline shape if the rock alone got harder to read.

observed occurrence = ψ (true occupancy) × p (detection probability)

Model 1. Sites you might have missed

Sixty-four sites. Each one is either truly occupied or not, set by the occupancy slider — you can reveal the ground truth, which a real field crew never gets to do. Each site is then surveyed some number of times; on each survey, an occupied site is detected with probability p, and an empty site is never falsely detected. Watch the naive "ever detected" occupancy chase the true occupancy as you add more surveys.

true occupancy ψ: 0.70naive "ever detected": –expected naive rate: –gap: –

Expected naive rate = ψ × (1 − (1−p)surveys). A single survey (surveys = 1) reduces to the plain ψ·p from the equation above. This is the club's own simulation of the mechanism in MacKenzie et al. (2002); it is not the 2025 paper's fitted model.

Model 2. The Cretaceous version

Same equation, run across the last 18 million years of the Cretaceous. Pick how detection probability behaves as rock exposure changes toward the boundary, and separately pick whether true occupancy is flat (the 2025 paper's finding) or actually declining (the rival hypothesis it argues against). Watch what the "apparent record" — the only thing anyone can actually dig up and count — does in each combination.

true occupancy ψ(t): –detection p(t): –apparent record: –
Drag the time slider from 84 Ma toward 66 Ma and watch the apparent record fall even while true occupancy stays flat.

Curves are the club's own illustrative shapes, chosen to match the qualitative pattern described in the paper and its coverage — a raw-record peak near 75 Ma and a decline into the final ~9 Myr — not digitised data from the paper's own figures.

Model 3. How many empty surveys before you believe "absent"?

One more piece of occupancy-model logic worth having in hand: if a site is truly occupied with detection probability p per visit, the chance of a false "empty" reading after n independent visits is (1−p)n. Set your own detection rate and see how many repeat looks it takes to get that chance below 5%.

Chance of a false "empty" after 1 survey
70.0%
Chance after 3 surveys
34.3%
Chance after 5 surveys
16.8%
Surveys needed to fall under 5%
9

A rock outcrop cannot be "resurveyed" the way a field ecologist revisits a pond: once it erodes away or a road gets built over it, that site's remaining chances to ever register a detection drop to zero. This is the structural reason a fossil record's effective survey count keeps falling even while researchers keep working just as hard.

What to take from pushing the sliders

The two big lessons sit in the same place both times. First, a single imperfect look at anything systematically undercounts it, and the undercount gets worse exactly where detection is worst — which is precisely where naive readers are most tempted to read "decline." Second, the fix is not more enthusiasm. It is modelling the detection process itself, with real covariates, so that a falling apparent count can be assigned to the right cause before anyone writes the animal's obituary.