Thesis Statement Worksheet

Copy to notes/scoping.md. Fill this in during weeks 1–4 and bring it to your advisor as the one-page problem statement. Expect to revise it repeatedly — that is the point of writing it early.

---

Part 1: Choose your angle

See guide/01-scoping.md for the full comparison. Answer the three deciding questions first, because two of them are hard constraints rather than preferences.

What hardware do I have exclusive, extended access to? (Be specific and honest. Four-hour slots on a shared cluster will not support careful latency measurement.)

How comfortable am I with training, validating, and debugging models? (Angle A punishes shallow ML skill.)

How much risk can my timeline tolerate? (A hard 6-month deadline argues for angle C or a narrow angle B.)

My angle: (A: AI for OS / B: OS for AI / C: Measurement)

Because: (two or three sentences — bring reasoning to your advisor, not an open question)

The main risk of this angle for me specifically, and how I mitigate it:

Part 2: The problem

Who has this problem?

What does it cost them? (a number if you can get one — a small motivating measurement of your own is very effective here and in your introduction)

Why is the obvious solution insufficient? (This is the paragraph that separates research from engineering. If you cannot answer it, the reader assumes you just did not think of it.)

Why now? (What changed — new hardware, new workloads, a newly practical technique?)

Part 3: The insight

One or two sentences. The key idea, before any mechanism. If a reader remembers one sentence from your thesis, this is it.

Part 4: The thesis statement

Formula:

By [mechanism], [system] achieves [quantified improvement] in [metric] on [workload class] compared to [named baseline], at a cost of [honest overhead].

Draft:

Now test it. Every "no" is something to fix.

• Could an experiment prove this false?
• Is the improvement quantified?
• Is the baseline named specifically?
• Is the workload class specified?
• Is the cost stated as part of the claim, not as a caveat?
• Is it one idea, not two?

Revision history (keep this — if it changes materially more than twice, see failure mode 2):

Part 5: Decomposed claims

Two or three claims that together support the thesis statement. If you cannot name the experiment that would test a claim, the claim is not concrete enough yet.

• C1:
• C2:
• C3:

Part 6: The gap

Which kind of gap am I claiming? (unexplored regime / unvalidated assumption / unmeasured cost)

The gap, in one paragraph:

The three closest papers, and how mine differs from each:

1. citekey —
2. citekey —
3. citekey —

Did I spend a deliberate afternoon trying to kill my own idea? (Search with other communities' vocabulary, check the last two years of the major venues, follow citations forward from your core papers. Finding the killer paper in week two is a good day.)

Date I did this: Result:

Part 7: Scope discipline

Cut list — good ideas I am explicitly not doing. Add to this whenever a new idea arrives; it also seeds your future work section for free.

Fallback: if the main idea does not work, what do I submit?

Part 8: The week-8 target

What must be working by the end of Phase 1?

• Baseline I must be able to measure:
• Benchmarks running with one command:
• Variance I must characterize:
• Machine access I must secure (and when I requested it):