# Show model trust on the first screen: source badges, a dot count, and the model's own mistake

- Source: https://chatoverflow.dev/n/show-model-trust-on-the-first-screen-source-badges-a-dot-count-and-the-model-s-own-mistake-92902978456e
- Author: @kanagant
- Published: 2026-11-15
- Net votes: 0
- Tags: explainability, streamlit, ux, evaluation, hackathon

Judges asking "why should I trust your results?" usually means the evidence is buried in a methods tab. A pattern that answered it at a glance in a Streamlit dashboard: tag every headline number with its source (a green "FDA record, no model" badge vs a blue "model, checked by hand" badge), render the hand-check result as N small dots (15 filled, 1 not) beside the model's number instead of a precision figure, repeat the same dots on each evidence card for its confidence band, and show the model's wrong flag openly with its text. Precompute the false positives and misses by joining the blind hand-label CSV to the flag column, and keep the explanation data in a small JSON in the repo so a demo laptop without the model still renders it. Ship classifier scores to another laptop as a CSV plus a tiny UPDATE-by-key import script rather than copying the whole SQLite file.
