3D Model Evaluation Simulator

Score field

Samples 900

Class separation 0.55

Score spread (noise) 0.16

Positive class prior 0.35

Sweep plane

Sweep speed 0.30

Threshold t

–

Pass

–

TP / FP

–

FN / TN

–

TPR / FPR

–

AUC-ROC

–

How it works

Each glowing sphere is one sample, placed along the depth axis by its classifier score — negatives (amber) and positives (cyan) drawn from two overlapping distributions. The translucent decision plane sweeps from high scores to low; every sample it has passed is now "predicted positive." As the plane sweeps, an orange trace grows on the ROC wall behind it, plotting (FPR, TPR) at that instant — one continuous sweep from (0,0) to (1,1), exactly as the ROC curve is defined.

plane at t → predict positive where score ≥ t
trace point = (FPR(t), TPR(t))
AUC = Σ trapezoid area swept so far

Sweep plane — orbit the scene while it runs; the plane's position IS the threshold, so watching it cross a cluster of points is watching that cluster flip from "predicted negative" to "predicted positive" in real time.
ROC wall trace — accumulates only in the direction the plane is currently moving; hit Reset trace to clear it and watch a fresh sweep build the curve from scratch.
Class separation / spread — a cleanly separated cloud produces a trace that shoots straight up the TPR axis before bending right; heavy overlap produces a trace hugging the diagonal.
Sweep speed — controls how fast the plane crosses the score axis; the underlying threshold, confusion counts and AUC integral are unaffected — this is purely playback speed.

Real-world relevance: an evaluation dashboard that only reports "AUC = 0.91" hides this whole sweep — the plane view makes explicit that AUC is an integral over every possible operating point, not a property of any single threshold choice.

Negative sample Positive sample Decision plane

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.