Setup
From Questions to Commitments
2 min
You have a research question from Breakout 1. Right now it's aspirational — it describes what you'd like to know. Your job now: turn it into something you could actually execute. That means stating your assumptions out loud, drawing boundaries, and defining the smallest experiment that would tell you if you're on the right track.
- Students often resist scoping down — it feels like giving up ambition. Reframe: "Scoping is not shrinking. It's choosing what to prove first."
- The MVP is not a product. It's the smallest experiment that would give you evidence for or against your research direction.
Phase 1
Surface Your Assumptions
8 min
Every research question hides assumptions. If you don't name them, they become invisible risks. Go through your research question and list everything you're taking for granted.
Assumption Audit
Research question
(from Breakout 1)
| Assumption | What if it's wrong? | How would you check? |
|---|---|---|
- Push for the "what if it's wrong" column — that's where the learning happens. If an assumption being wrong wouldn't change anything, it's not a real assumption.
- If a pair lists fewer than 3 assumptions, they're not looking hard enough. Ask: "What about the data? The labels? The evaluation?"
Phase 2
Define Your MVP
8 min
Now scope it down. What's the smallest experiment you could run — in two weeks, with the data you have — that would tell you whether this direction is worth pursuing? Not the full system. Not the paper. The first thing you'd try.
MVP Definition
In scope
Out of scope
Data subset
"Which slice of the data? How much? Why that slice?"
Simplest method
"What's the dumbest thing that could work?"
Success signal
"What result would make you keep going? What would make you stop?"
Timeline
"Can you do this in 2 weeks? If not, scope smaller."
- "What's the dumbest thing that could work?" is the key question. If K-means on TF-IDF already solves the problem, there's no research question.
- Good MVPs sound boring: "Run K-means on 500 conversations with 3 different embeddings and have 2 people judge if the clusters make sense." That's two weeks of work and it tells you a lot.
- If a pair's MVP is "build a deep learning pipeline" — push back: "What do you need to know before you build that?"
Synthesis
Share
2 min
Let's hear from 2–3 pairs. Read us your MVP — just the "simplest method" and the "success signal." We'll see if the rest of the room thinks it's actually achievable in two weeks.
- Pick pairs whose MVPs are at different ambition levels — one that's well-scoped and one that's still too big. Let the room see the contrast.
- Ask the class: "Could they do this in two weeks? What would you cut?"