The Map: Six Phases and the Gates Between Them

This is the shape of a 6–12 month monograph thesis. The week numbers below assume a 40-week schedule; compress or stretch them proportionally, but keep the ratios, because the ratios are where people go wrong.

The single most common scheduling error in systems research is treating writing as something that happens after the research. It isn't a phase at the end. Related work gets drafted in month two, while the papers are still fresh in your head. The design chapter gets drafted as you design, not remembered six months later.

The second most common error is under-budgeting infrastructure. In OS work, getting to your first trustworthy baseline number reliably takes longer than you think — kernel builds, benchmark harnesses, measurement plumbing, machine access, permissions. Budget generously and be suspicious of any plan where experiments start in week two.

The phases

Week   0        4        8       12       16       20       24       28       32       36    40
       |--------|--------|--------|--------|--------|--------|--------|--------|--------|-----|
P0  Scope  ████████
P1  Infra + baseline    ████████████████
P2  Core work                   ████████████████████████████
P3  Evaluation                                   ████████████████████
P4  Writing            ░░░░░░░░░░░░░░░░░░░░░░░░░░████████████████████████
P5  Revision + defense                                             ████████████████████████
                                                    ↑
                                              RESULTS FREEZE (week 30)
       ░░░ = incremental drafting alongside research    ███ = primary focus

Phase 0 — Scope lock (weeks 1–4)

Turn "I am interested in X" into "I claim Y, and here is how I would know if I were wrong." Read guide/01-scoping.md.

Also do the administrative work from the README this month: formatting rules, deadlines, machine access requests, ethics or IP paperwork if your work touches either. Machine access in particular can have a multi-week lead time in university clusters, and discovering that in month four is painful.

Gate — do not leave Phase 0 until: you have a written one-page problem statement with a falsifiable thesis statement and two or three specific claims, your advisor has read it and agreed, and you can name the strongest three papers your work must be distinguished from.

Phase 1 — Infrastructure and baselines (weeks 3–8)

Build the measurement apparatus and characterize the existing system before you build anything of your own. You need a testbed you can control, benchmarks that run with one command, results that land in a structured directory, and baseline numbers you trust enough to defend.

This phase is where you discover the unglamorous facts that shape the whole thesis: that the workload you assumed is I/O-bound is actually lock-bound, that variance between runs is 15% so your hoped-for 8% improvement is unmeasurable without more care, that the "default" configuration everyone compares against is badly tuned. Better to learn all of this now.

Read guide/06-evaluation-playbook.md during this phase, not later. It describes the apparatus you are building.

Gate: you can produce a baseline measurement with one command, on a controlled machine, and its run-to-run variance is small enough that the effect size you are hoping for would be detectable. If your variance exceeds your expected effect, fix the measurement before proceeding — no amount of later cleverness recovers from this.

Phase 2 — Core work (weeks 8–28)

Build the thing. Iterate. This is the longest phase and the one that needs the least advice, except this: keep a dated engineering log, and every two weeks ask yourself whether the thing you are now building is still the thing your thesis statement claims. Drift here is silent and expensive.

Draft the design chapter as you go. The rationale for a decision — including the alternatives you rejected and why — is vivid the week you make it and nearly unrecoverable four months later. Those rejected alternatives are what make a design chapter read as research rather than as documentation.

Gate: an end-to-end working system, however rough, that produces a result you can compare against the Phase 1 baseline. Not feature-complete. Working.

Phase 3 — Evaluation (weeks 16–30)

Run the experiments your claims-to-experiments matrix (from guide/03-research-design.md) says you need. This overlaps Phase 2 heavily and intentionally: early partial evaluation tells you what to fix while there is still time to fix it.

Include the experiments you are afraid of — the adversarial workload, the case where your mechanism should lose. You want to find those yourself, in month six, rather than have your committee find them in month eleven.

Gate — the results freeze. Pick a date, roughly week 30, after which you run no new experiments except ones that fix an error or fill a hole a reader identified. Write it down and tell your advisor. Without a freeze, "one more experiment" consumes the writing budget, and a thesis with good results and a rushed write-up scores worse than one with adequate results and a clear argument.

Phase 4 — Writing (weeks 24–38)

Assemble and write the document. Read guide/04-thesis-structure.md for what goes where and guide/05-writing-process.md for how to actually produce the pages.

Budget generously for your advisor's turnaround. Two to three weeks per chapter is normal and you cannot compress it by wishing. If you have four chapters needing review and your advisor takes two weeks each, that is eight weeks of latency you must plan around by sending chapters as they finish rather than in one batch at the end.

Gate: a complete draft with no placeholder sections. Every figure present, every citation resolved, conclusion written.

Phase 5 — Revision and defense (weeks 32–40)

Revision passes, committee feedback, defense preparation, submission logistics. Read guide/07-defense-and-submission.md.

Gate: submitted, defended, corrections filed.

Milestones to put in your calendar

Dates beat intentions. Set these as real calendar events with reminders.

The mid-point review

At week 24, sit down with your advisor and answer three questions in writing.

Is the thesis statement still what my results actually support? Often it is not, and the honest move is to narrow the claim to fit the evidence rather than stretch the evidence to fit the claim. Narrowing is not failure — an overreaching claim is the most common thing committees attack.

What is the weakest part of my evidence, and can I still strengthen it? At week 24 you have time. At week 34 you do not.

If everything from here fails, what do I submit? You should have an answer. See the fallback discussion in guide/01-scoping.md. A thesis that honestly reports "this approach does not work, here is rigorous evidence for why, and here is what that implies" is a legitimate and defensible thesis. One that reports nothing because the student was still hoping is not.