# Muse Wonder Side-by-Side Model Comparison App

## PURPOSE
This project rebuilds the core workflow of the video "Muse Wonder FULL from Meta | First Impressions": showing one prompt and stacking the outputs of two models beside each other so a human can judge which is better. In the transcript the reviewer repeatedly places "Muse Wonder" next to "Muse Wonder FULL", then next to Opus, Sonet, Astra and Argon, and narrates specific artefacts ("a bear floating midair", "birds swimming", "fish slamming their heads into the walls"). The app captures that exact pairing plus a free-text artefact note field, so a student can turn a casual impression into a structured verdict.

## WHY THIS MATTERS
The central lesson is that public benchmark rankings and hands-on testing disagree. The non-full model sat 17th on the code arena; the full max-reasoning version jumped to 6th, yet the reviewer still says "it's not blowing me away". A comparison harness is how you reproduce that judgement instead of trusting a leaderboard number.

## HOW TO BUILD
1. Build a Gradio `Blocks` layout with a shared prompt box and two output columns.
2. Wrap the model call in a `respond(model_id, prompt)` function.
3. Store one verdict row per (prompt, model) pair in a small JSONL log.
4. Render a summary table of pairwise wins.

```python
import gradio as gr

with gr.Blocks(title="Muse Wonder Comparison") as demo:
    prompt = gr.Textbox(label="Prompt", lines=2)
    with gr.Row():
        model_a = gr.Dropdown(["muse-wonder", "muse-wonder-full"], value="muse-wonder")
        model_b = gr.Dropdown(["muse-wonder", "muse-wonder-full"], value="muse-wonder-full")
    with gr.Row():
        out_a = gr.Textbox(label="Output A")
        out_b = gr.Textbox(label="Output B")
    run = gr.Button("Compare")
    run.click(compare, [prompt, model_a, model_b], [out_a, out_b])
demo.launch()
```

## WHEN TO USE IN PRODUCTION
Use it for internal model bake-offs before a vendor decision, for regression checks after a provider updates a model, or for collecting human preference labels. Do NOT use it as an automated evaluator — it is a human-in-the-loop tool; without a grader it says nothing about quality on its own.

## HOW IT HELPS STUDENTS
Students learn Gradio event wiring, provider abstraction behind one interface, deterministic stubbing for offline runs, and how to convert subjective impressions into a logged, reviewable dataset — all portable to real eval harnesses.

## STARTER CODE
See `starter/main.py` for the runnable skeleton: a `respond()` helper with a deterministic stub, the Gradio layout, and TODO markers for wiring a real model API.
