一份研究陈述只有在它承认哪些还没解决的时候才有用。下面三个项目按「我敢在公开场合为其中多少辩护」排序。
A research statement is only useful if it admits what is still open. The three projects below are ordered by how much of them I would defend in public.
把原本只能手工量的东西从田里取出来——根系、冠层、一个生长季里的变化——而不是让一个人坐下来一株株数。
Getting measurements out of a field that would otherwise be taken by hand — root systems, canopy, growth over a season — without asking a person to sit and count.
在田间真实存在的条件下做检测与分割：叶片互相遮挡、光照一直在变，照片也永远不会像基准数据集那样干净。
Detection and segmentation under the conditions a field actually has: overlapping leaves, changing light, and photographs that will never be as clean as a benchmark.
一个被报告出来的提升，在固定评估协议之后还能剩下多少；以及一个小而吵的样本能诚实支撑多少信心。
How much of a reported improvement survives a fixed evaluation protocol, and how much confidence a small, noisy sample can honestly support.
从消失元的结构反过来刻画有限群；以及追问一个融合系统在什么条件下会从它的来源群继承超可解性。
Characterising finite groups from the structure of their vanishing elements, and asking when a fusion system inherits supersolvability from the group it came from.
让别人能真的把方法跑起来的界面与流程。一个除了自己没人跑得起来的方法，还不算产出了结果。
Interfaces and pipelines that let the next person run the method. A method nobody else can run has not produced a result yet.
在不逐帧人工标注的前提下，低成本采集的图像里能还原出多少细根结构？
How much fine-root structure can be recovered from low-cost captures without manual annotation of every frame?
带显式遮挡模型的多帧融合，以标注方差为基线来评估，而不是拿一份手工标注当标准答案。
Multi-frame fusion with an explicit occlusion model, evaluated against an annotation-variance baseline rather than a single hand-labelled set.
在 F_S(G)-融合系统中，哪些结构性条件会强制超可解性？
Which structural conditions force supersolvability in F_S(G)-fusion systems?
群论论证；当前这一版在尝试削弱主定理的假设条件。
Group-theoretic arguments; the working revision is an attempt to weaken the hypotheses of the main theorem.
固定评估协议之后，已发表的提升还能剩下多少？
How much of a reported gain survives a fixed evaluation protocol?
把一批已发表的配置放到同一套划分和同一套预处理下重跑，报告前后的相对排名。
Re-ran a set of published configurations under one split and one preprocessing pipeline; reported the relative ranking before and after.
这里每个项目最后都会汇到同一个第一步：先确定测量本身给误差贡献了多少。这个数字出来之前，方法之间的比较只是噪声之间的比较。
Every project here converges on the same first step: establish what the measurement itself contributes to the error. Until that number exists, comparisons between methods are comparisons between noise.