菜单
Use when the research task is in neuroscience — neural recording, decoding and brain-model comparison — at study design, analysis or writing. Two-sided comparator ladder, the negative-control representation panel, and per-unit predictions
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/tangxiangru/AutoR --skill neuroscience-comparator-ladder-and-per-unit-predictions命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Use when writing results and discussion, and when a task or a reviewer asks why an effect happens rather than whether it does. Covers the difference between reporting an effect and accounting for it, and what a mechanism claim needs behind it.
Use when the supplied data looks synthetic, degraded, incomplete or wrong, and whenever you are tempted to reframe the study around what you found about the inputs, the harness or the evaluation. Covers what to do with a real data problem without losing the study.
Use when planning figures, at study design and again before writing. Covers the figures a paper in this field is expected to contain, why an original figure does not substitute for a standard one, and how to decide what to draw first.
基于 SOC 职业分类
正在显示 SKILL.md
| name | neuroscience-comparator-ladder-and-per-unit-predictions |
| description | Use when the research task is in neuroscience — neural recording, decoding and brain-model comparison — at study design, analysis or writing. Two-sided comparator ladder, the negative-control representation panel, and per-unit predictions |
Three deliverables a computational-neuroscience claim must carry; plan them before fitting.
A two-sided comparator ladder. Horizontally, a bank of named alternative methods spanning the families in use (supervised, correlation based, variance based, single-modality). Vertically, variants of your own model that each delete one information source your thesis says is necessary - structure or connectivity, task optimisation, temporal order, one modality - plus a granularity sweep that coarsens the entity taxonomy (fine type, family, broad class) until performance collapses. Same metrics, same split, every rung.
The paired representation panel. Show the low-dimensional projection twice on identical axes and colouring: once under the untreated, shuffled or degraded condition and once under the treated one, quantified with the same gap metric in both. The deliberately poor control panel is a required deliverable, not something the good panel excuses.
Per-unit publication. Give the model's per-unit quantity for the whole population as a ranked table or figure, partitioned into units where an independent measurement exists (report agreement as k of n) and units where the model issues an untested prediction, labelled as novel predictions, with the coverage fraction stated. Pair it with a mechanism established by intervention - property P of upstream unit A sets property Q of downstream unit B, shown by silencing or re-weighting A - closing with a prediction an experimentalist could test.
Each must appear as a labelled figure with its numbers in the panel or caption. A quantity computed into a results file but never plotted counts as not done.
Three tasks each demand the comparator ladder, the paired projection with its deliberately-bad control panel, and the intervention-based mechanism statement; three demand the per-unit table split into validated versus novel predictions with a stated coverage fraction. Bare runs ran textbook single-input ablations instead of the variant grid, never ran a taxonomy-coarsening sweep, and never drew the negative-condition panel. The closing rule targets the largest measured single loss in the discipline: per-feature attributions written to a diagnostics JSON, never plotted and never named in the report, taking 0.60 of one task's weight to zero.