Conservative Wording Guard (Over-Claim Protection)

A reference for conservative wording in Chinese degree theses. The goal is not to make statements weak, but to state evidence strength precisely: strong evidence permits strong wording; weak evidence requires weak wording. Blind-review experts are highly alert to over-claiming, so proactive calibration is easier than revising after questioning.

Boundary with “Is the Evidence Sufficient?”

This file governs how strong the wording should be once evidence strength is known: which verb or qualifier keeps the sentence within the evidence. First determine whether figures/tables/metrics/citations actually support the claim, then use this file to choose wording. Substance takes priority when it conflicts with wording.

Certainty Ladder (Strong to Weak)

证明 / 表明（强）                 ← 干预实验（消融/受控对比）
  ↓
揭示 / 发现 / 识别出               ← 强效应，多方法或可复现
  ↓
表明 / 提示                       ← 显著但单一方法
  ↓
支持 / 与……一致                  ← 趋势性，与前人一致
  ↓
可能表明 / 或许提示 / 似乎         ← 边缘显著或预测性
  ↓
暗示 / 倾向于                     ← 极弱信号或假说

Match the rung to the evidence; do not climb above what the data can reach.

Replacement Tables

1. Causality (Most Common Over-Claim: Calling Correlation Causation)

❌ Over-Claim	✅ Conservative Wording
由……导致 / 引起	与……相关 / 关联
驱动 / 决定	影响 / 与……相关
是……的根本原因	与……有关 / 可能与……有关
证明了	表明 / 提供了……的证据
造成了	伴随出现 / 与……同时出现

Use causal wording only for controlled interventions (ablation, randomized grouping, A/B comparison), instrumental-variable designs, or reproduction of an established mechanism. Otherwise use correlational wording.

2. First/Unique Claims (Reviewers Will Search Immediately)

❌ Over-Claim	✅ Conservative Wording
首次 / 第一个	据我们所知，首次 / 最早的工作之一
新颖的（自我标榜）	Directly state what is new; delete the “novel” label
前所未有的	显著的 / 值得注意的
此前未知的	此前研究不充分的

3. Universality (One Scenario Cannot Support Every Scenario)

❌ Over-Claim	✅ Conservative Wording
总是 / 永远 / 从不	通常 / 很少
在所有情况下	在所研究的情形中
普遍地	在所评测的基准上
任意数据集	所采样的数据集

4. Effect Size (“Significant/Large” Without Numbers)

❌ Over-Claim	✅ Conservative Wording
大幅提升	误差降低了 X%
显著的效应	β = X.XX（95% CI：…）
明显改善	从 X 提升到 Y（p = …）
高度显著	p < 1 × 10⁻¹⁰
鲁棒 / 稳健	在 N 次独立运行中一致 / 在[扰动]下稳定

If the number already tells the story, remove the adjective and let the number speak.

5. Time/Inference Order (Inferring Historical Causality from Contemporary Data)

❌ Over-Claim	✅ Conservative Wording
X 驱动了该变化	该变化与 X 一致
发生在 T 时刻	估计约为 T（置信区间：…）
从 A 迁移到 B	数据与 A→B 的路径一致

6. Application Prospects (Downstream Use Not Demonstrated Here)

❌ Over-Claim	✅ Conservative Wording
将带来变革	对……具有潜在意义
将被广泛使用	可能有助于 / 可为……提供参考
解决了 X 问题	处理了 X 的一个方面
可直接落地部署	为[场景]提供了候选方法

7. Comparison (Disparaging Prior Work)

❌ Over-Claim	✅ Conservative Wording
前人方法未能……	前人方法受限于……
优于所有已有方法	与[具体方法]相比具有优势
终结了长期争论	为该争论的一方补充了证据

Frequent Trap Sentences

Trap	Safe Alternative
“本文结果证明了 X。”（X 是因果）	“本文结果与 X 一致。”
“这是首个……的工作。”	“据我们所知，是最早……的工作之一。”
“X 在 Y 中起关键作用。”	“X 与 Y 有关 / 可能对 Y 有贡献。”
“这些发现对……具有重要意义。”	“这些发现为进一步研究……提供了基础。”
“X 是 Y 的关键驱动因素。”	“X 与 Y 相关。”
“强烈支持”	“与……一致 / 提供了与……一致的证据”

Reverse Calibration: When Not to Be Conservative

Weak evidence requires caution, but cautious wording for strong evidence becomes timid. Use strong wording when:

• a controlled intervention (ablation / randomized control / A-B) yields a causal result -> use “证明”;
• multiple methods/datasets/random seeds reproduce the result -> use “稳健” and state the evidence;
• an established mechanism is reproduced -> “确认 / 验证” is appropriate;
• a large effect has strong statistics -> use strong wording with the number.

Self-Check After Each Paragraph

• [ ] Used “首次/新颖”? Was the literature actually searched? If not, add “据我们所知”
• [ ] Used “导致/驱动/决定”? Is there an intervention? If not, change to “与……相关”
• [ ] Used “所有/总是/普遍”? Is the scope limited to the actual study?
• [ ] Used “显著/大幅/明显”? Is a number attached?
• [ ] Listed an application not demonstrated here? Add “可能/或许”
• [ ] Disparaged prior work? Use “受限于” instead of “未能/失败”

Script Support

deai_check.py emits a [Script] LOW trace for a small set of unambiguous over-claim phrases in causality/first/universality/application categories and points back here. The script catches only obvious cases; the tables above cover judgments the script cannot make.