|
Download SKILL.md from bhxdianzhang/CheckSectionLogicConsistency: direct link, hf CLI and curl.
- Browser
- Download file 4.38 kB
-
https://huggingface.co/datasets/bhxdianzhang/CheckSectionLogicConsistency/resolve/ae08bfb2dacd5606ae044df435bc32da9feb483c/SKILL.md
- Command line
-
hf download hf://datasets/bhxdianzhang/CheckSectionLogicConsistency@ae08bfb2dacd5606ae044df435bc32da9feb483c/SKILL.md
-
curl -L -o SKILL.md https://huggingface.co/datasets/bhxdianzhang/CheckSectionLogicConsistency/resolve/ae08bfb2dacd5606ae044df435bc32da9feb483c/SKILL.md
4.38 kB
metadata
name: check-section-logic-consistency
description: >-
Check whether abstract, introduction, method, experiments, and conclusion use
consistent terminology, claims, assumptions, and closure structure. Use when a
paper feels internally inconsistent, has claim drift, or different sections
appear to promise different contributions. Output is a consistency report and
repair plan.
metadata:
owner_model: ParadoxGPT-Checker-4B
skill_family: checker
version: '0.1'
output_type: diagnosis
default_language: zh
trigger_keywords:
- logic consistency
- section consistency
- claim drift
- terminology
- contradiction
- 前后一致
- 逻辑
required_inputs:
- paper_sections
optional_inputs:
- claim_list
- terminology_list
- reviewer_concerns
handoff_to:
- check-claim-evidence-alignment
- build-intro-argument-chain
- rewrite-abstract-with-commitments
Check Section Logic Consistency
Purpose
检查论文各章节是否讲的是同一个故事:abstract 的承诺、intro 的 gap、method 的设计、experiment 的证据、conclusion 的总结是否一致。这个 skill 关注跨章节漂移,而不是单点 claim 证据。
When to Use
- 读起来感觉 abstract、intro、method、experiment 各讲各的。
- 术语在不同章节含义变化。
- 贡献从 intro 到 conclusion 被放大或换了目标。
Do Not Use
- 只检查 claim 是否有证据 → 用
check-claim-evidence-alignment。 - 需要重建 introduction 论证链 → 用
build-intro-argument-chain。 - 只改 abstract → 用
rewrite-abstract-with-commitments。
Required Inputs
paper_sections: abstract、introduction、method、experiment、conclusion 的文本或摘要。claim_list: optional,已抽取 claims。terminology_list: optional,关键术语清单。
Output Contract
输出必须至少包含:
- terminology_inconsistencies - 术语漂移或定义不一致。
- claim_drift - claim 在章节之间变大、变小或换目标。
- missing_closure - 某章节提出但后文没闭合的承诺。
- contradictions - 明确矛盾或隐含冲突。
- repair_plan - 按优先级修复。
Procedure
- 抽取每个章节的主 claim 和关键术语。
- 对齐同一术语在不同章节的定义和作用。
- 检查 intro 的 gap 是否真的由 method 回应。
- 检查 experiments 是否测试 abstract/conclusion 的核心承诺。
- 输出 repair_plan,优先修会误导 reviewer 的 drift。
Quality Bar
一个好的输出必须:
- 指出具体章节间的不一致,而不是泛泛说逻辑不顺。
- 区分 terminology drift、claim drift 和 missing closure。
- repair_plan 能说明改哪一节最划算。
- 不把合理的逐步细化误判为矛盾。
一个差的输出:
- 只做全文摘要。
- 把所有表达差异都当 inconsistency。
- 不指出修复位置。
Failure Modes
- 过度严格:术语自然细化不等于漂移。
- 漏掉 conclusion 扩张:结论常把 claim 写大。
- 忽视方法目标:method 优化的目标和 intro 诉求不一致。
- 只看字面:同义改写不一定不一致,要看 claim 范围。
Handoff
- 如果漂移本质是证据不支撑 claim →
check-claim-evidence-alignment - 如果 intro 链条需要重建 →
build-intro-argument-chain - 如果 abstract 承诺与正文不一致 →
rewrite-abstract-with-commitments
Example
Input:
paper_sections:
abstract: claims scalable interaction for long-horizon reasoning
introduction: motivates memory failures
method: introduces interaction budget scheduler
experiments: tests task success, not memory consistency
conclusion: claims general agent memory improvement
Output:
terminology_inconsistencies:
- “memory” in intro means persistent state, but method only changes interaction scheduling.
claim_drift:
- abstract: interaction scaling.
- conclusion: agent memory improvement. This shifts the contribution target.
missing_closure:
- intro's memory failure motivation is not tested by experiments.
contradictions:
- no direct contradiction, but contribution framing drifts.
repair_plan:
1. Reframe intro around interaction budget, not memory.
2. If keeping memory framing, add memory consistency diagnostic.
3. Rewrite conclusion to match tested interaction claim.