Load prediction data
Select the four CSV files from your repo. Each file must have columns True Label and Class_0_Probability … Class_9_Probability. Everything runs locally — no data leaves your machine.
MNIST · before calibration
MNIST_PRED / bef_100.csv
click to select
MNIST · after calibration
MNIST_PRED / aft_100.csv
click to select
CIFAR-10 · before calibration
CIFAR_PRED / bef_100.csv
click to select
CIFAR-10 · after calibration
CIFAR_PRED / aft_100.csv
click to select
Files are read directly in your browser · nothing is sent anywhere
Auto-loads from MNIST_PRED/ & CIFAR_PRED/ when served via HTTP
On file:// use “Load folder”, paste a remote base URL and click “Load data”, or pick files individually
SL
Trust
Evidential Based Trust Assessment Framework · Subjective Logic
PhD Demo
Trust opinion
Global SL opinion ω = (belief, distrust, uncertainty) over the last checkpoint. Derived from calibration evidence across all classes — no model internals required.
Belief t
—
Evidence for trust
Distrust d
—
Evidence against
Uncertainty u
—
Vacuity
ECE
—
Calibration error
Accuracy
—
Top-1
ω composition
Belief
Distrust
Uncertainty
Reliability diagram
Per-class opinion
Calibration-based trust
Predicted probabilities are partitioned into M clusters. Per-cluster belief/distrust evidence is counted and fused cumulatively via the Subjective Logic cumulative belief fusion operator — without accessing model internals.
Chapter 8 · C3
10
Belief t
—
Distrust d
—
Uncertainty u
—
ECE
—
ω composition
Reliability diagram
Per-cluster opinion (stacked)
Effect of M on the trust opinion
How belief, distrust and uncertainty converge as the number of probability clusters M grows. Evaluated at the final epoch for before and after calibration.
Chapter 8 · Figure 2
1000
Before calibration — opinion vs M
After calibration — opinion vs M
Dynamic per-prediction assessment
At inference time, each prediction's confidence maps to the nearest cluster representative. The pre-computed cluster opinion becomes that prediction's trust score — enabling selective abstention.
Chapter 8 · Figure 3
10
Mean belief — correct
—
Mean belief — wrong
—
Separability Δ
—
Distribution of predicted confidence
Distribution of trust belief — correct vs incorrect
Mean SL opinion per confidence bin
Evaluate your model
Upload a predictions CSV or a raw dataset + ONNX model. The full SL calibration trust pipeline runs in-browser — no data leaves your device.
Predictions CSV
Model + Dataset
Data scope
Expected columns: True Label, Class_0_Probability, Class_1_Probability, … Class_N_Probability
⬆
Drop predictions CSV or click to browse
10