Topic selectiontopic-selection/proposals/README

Proposals

Full one-page problem statements for the two shortlisted options, in the shape guide/01-scoping.md demands: the problem and who has it, why the obvious solution fails, why now, a falsifiable thesis statement, decomposed claims, the gap in one paragraph naming the three closest papers, the fallback, and what must work by week eight.

Each is longer than a page because it carries the evidence you will need in your first advisor meeting. Compress to one page before sending. The advisor gets the argument; you keep the citations.

File	Topic	Lane	Kernel depth	Can it fail outright
`A-memory-hierarchy-citizenship.md`	What local inference costs the machine it runs on	Systems for AI, inverted	2 — eBPF	No
`B-khugepaged-collapse-creep.md`	Predicting collapse retention instead of thresholding fullness	AI for Systems	3 — core MM	Yes

B closes with a side-by-side comparison of the two.

These are not neutral documents

Both were written against one constraint profile: shared-cluster GPU access with real queue time, eBPF as the only kernel skill already in hand, an OS/kernel advisor, roughly twelve part-time months, and a publication at a decent venue as the goal. Two consequences are worth restating because they are counterintuitive.

The cluster is a liability, not an asset. Every topic that would use it needs exclusive, uncontended GPU time for tail-latency work, and guide/01-scoping.md says so plainly: "a shared cluster where you get four-hour slots will not support careful latency measurement." Both proposals therefore choose topics where a single machine is the correct platform.

Twelve part-time months is roughly six full-time-equivalent, which eliminates every twelve-month entry in ../01-candidate-topics.md regardless of merit, including the sched_ext option rated least crowded there.

If your constraints differ, re-derive the shortlist rather than inheriting it.

The property that matters most

Both proposals have a cheap week-two or week-eight experiment that can kill them, and that is deliberate rather than incidental. For A it is whether correctly configured cgroup v2 confinement already closes the gap. For B it is whether max_ptes_none=0 actually forgoes measurable TLB benefit. Run the kill test before Phase 2, not after — guide/01-scoping.md is explicit that finding the killer in week two is a good day and finding it in month eight is a catastrophe.

Before you commit to either

Re-verify crowdedness yourself. Both are dated 26 August 2026, and A has roughly a one-year shelf life because BPF OOM and the RAM-backed-swap reclaim fix are both in flight.
Do the reading in ../03-reading-list.md under the matching proposal heading — two days, first pass, reading for limitations rather than mechanisms.
Run the idea-killing search named at the end of each reading section, using the other community's vocabulary.
Take the compressed page to your advisor with a recommendation and reasoning, not as an open question.