Introspection | A knowledge manifold

CHROME BUILT-IN / GEMINI NANO

Chrome handles the model.

Gemini Nano runs through Chrome’s Prompt API. Prompts and generations stay on this device.

MODEL FILES: Not exposed
FIRST LOAD: Varies by device

MODEL CONFIDENCE

Model self-assessment

The model rates how well it knows an entity from 0–100. Chrome does not expose token probabilities.

GEMMA 4 E2B / WEBGPU

Open weights expose token probability.

Gemma runs in a dedicated browser worker. Its model files can remain in the browser cache.

MODEL FILES: 3.4 GB
OBSERVED LOAD: ≈4 min

MODEL CONFIDENCE

Mean log probability

Gemma scores every token in the generated type and description. We average their log probabilities, then exponentiate the mean to produce the 0–100 confidence score.

The cited study found this length-normalized sequence log probability was the most effective method tested for detecting hallucinated translations.

Guerreiro, Voita, and Martins, 2022