# Generation Quality & Cost Dashboard

## PURPOSE
A Streamlit + Plotly dashboard that turns the video's spoken evaluations into an explorable chart. Across "Muse Wonder FULL from Meta | First Impressions" the reviewer walks task families — 3D scenes, HTML video clips (Great Wall of China, Taj Mahal, a Trojan horse on a vase), ramen-shop / chocolatier / ski-resort websites, and a Grand Prix data dashboard — and ranks models within each: Opus strongest, Astra close behind, Sonet vibrant but over the top, Argon heavy, Muse Wonder cartoonish, Muse Wonder FULL mid-pack. This project encodes those per-category scores and generation costs into interactive plots.

## WHY THIS MATTERS
The whole video is one long argument that rankings and inputs should be inspected together, not summarised to a single number. The reviewer flags that Muse Wonder looks "cartoonish", that Sonet is "so artistic" but may not cope with complex UX, and that cost is a caveat on everything ("$11", "$15, nearly 16", the full model unpriced). A dashboard makes category-by-category trade-offs and the cost overlay visible at a glance.

## HOW TO BUILD
1. Store scores in a long-format Pandas frame: `model, category, score, cost_usd`.
2. Build a Plotly grouped bar chart with models on x and category as colour.
3. Add a cost scatter (score vs cost) so the frontier is obvious.
4. Use a `st.sidebar.multiselect` to filter models and categories.
5. Cache the data load with `@st.cache_data`.

```python
fig = px.bar(df, x="model", y="score", color="category", barmode="group",
             title="Generation quality by task family")
st.plotly_chart(fig, use_container_width=True)

scatter = px.scatter(df, x="cost_usd", y="score", color="model",
                     hover_data=["category"], title="Quality vs generation cost")
st.plotly_chart(scatter, use_container_width=True)
```

## WHEN TO USE IN PRODUCTION
Use it for model-selection reviews, steering-committee dashboards, or a live "capability matrix" page when a vendor ships a new variant. Avoid it for precise quantitative benchmarking — the scores here are qualitative expert judgements; and never present it as a single-model winner, since the video's point is that the answer is category-dependent.

## HOW IT HELPS STUDENTS
Students get hands-on Streamlit layout, Plotly figure construction, wide-to-long reshaping, caching, and — importantly — how to visualise human judgement data honestly with uncertainty rather than overstating a metric.

## STARTER CODE
`starter/main.py` ships a transcript-seeded DataFrame (categories and relative scores derived from the review), the two Plotly figures, and TODO markers for adding confidence bands and an editable scoring form.
