Answer consistency analysis

Measure when answers stop agreeing.

Compare sampled responses with bidirectional entailment. See how they cluster, how uncertain the set is, and whether it crosses your guardrail.

Answer samples

One response per line

Responses from your model

3 to 5 samples · 4,000 characters each 4 / 5

Semantic cluster threshold

0.80

Entropy limit

0.50

Minimum top-cluster margin

0.20

Model loads the first time you evaluate.

Inference runs locally in your browser. Only model files are downloaded from the Hugging Face Hub; your sample text is not sent to a server.

Not loaded

Run an evaluation to load the model, cluster the samples, and calculate the uncertainty metrics.