# Gemini 4 Argon Field-Report Dashboard

## PURPOSE
The video is structured as a field report: it explicitly promises "the good, the bad, and
the long" and says it will show less of the official demos and more of what is out there in
the wild. That structure is a dataset waiting to be visualised. This Streamlit and Plotly
dashboard takes survey rows — model, criterion, sentiment, score, source, date — and renders
the four claims the clip actually makes: conversation quality and instruction-following are
up and "less robotic," long-conversation memory holds, pricing is aggressive at the intro
rate, and coding/access is the weak spot.

## WHY THIS MATTERS
Model releases arrive wrapped in cherry-picked demos. The reviewer's own methodology — pull
a production copy off the shelf rather than accept the vendor's review unit — is the
correction. A dashboard that separates vendor claims from independent reports, and that
shows the negative column as loudly as the positive one, is what a careful practitioner
actually needs on release day.

## HOW TO BUILD
1. `pip install -r requirements.txt` (Streamlit, Plotly, Pandas).
2. Load `reports.csv` with columns `model, criterion, sentiment, score, source, date`.
   Ship a template file with the transcript's own quotes as the seed rows.
3. Build three views in tabs: a diverging bar of mean score per criterion split by
   sentiment, a line of sentiment over days since release, and a pricing panel plotting
   $/1M output tokens with the introductory rate flagged.
4. Add a model filter in the sidebar and a raw-rows expander so numbers are auditable.
5. Cache the CSV load with `@st.cache_data` and never mutate the source frame.

```python
import streamlit as st, plotly.express as px, pandas as pd
df = st.cache_data(pd.read_csv)("reports.csv")
fig = px.bar(df, x="criterion", y="score", color="sentiment", barmode="group",
             title="Field reports by criterion")
st.plotly_chart(fig, use_container_width=True)
```

## WHEN TO USE IN PRODUCTION
Use it for release-postmortems, internal model scorecards, and evaluation reporting where
humans are hand-labelling quality. Do NOT use it for automated benchmark dashboards — those
belong in CI with promptfoo-style harnesses. Skip it when you have fewer than a few dozen
rows; a table is more honest than a chart of five points.

## HOW IT HELPS STUDENTS
They practise tidy-data design, sentiment aggregation, honest axis and label choices, and
building an app that shows inconvenient results rather than hiding them.

## STARTER CODE
`starter/main.py` ships an embedded template dataset derived from the transcript's stated
claims, with `# TODO` markers for wiring a real survey CSV and live pricing.
