Skip to content

Evidence and Integrity

Connor edited this page Jul 3, 2026 · 2 revisions

Evidence and Integrity

This skill is designed to reduce the most common failure modes in academic writing: invented papers, inflated SOTA claims, fake results, unsupported claims, and invented local requirements.

Non-Negotiable Rules

  1. No fabricated papers.
  2. No fabricated SOTA.
  3. No fabricated results.
  4. No invented school, college, advisor, template, rubric, defense, or workload requirements.
  5. Prefer primary sources.
  6. Map every strong claim to evidence.
  7. Reduce or qualify claims instead of overclaiming.
  8. Load only needed modules.
  9. Keep external skills low-coupled.
  10. Require human review for high-stakes decisions.

Evidence Types

Evidence Type Examples
Literature Verified papers, official datasets, official code, venue pages
Experiments Completed runs, metrics, baselines, ablations, logs
Implementation Code, tests, reproducibility notes, configuration
Requirements Official school templates, advisor documents, venue instructions
Proof or reasoning Formal proof, derivation, design rationale, failure analysis

Result Status

When writing experimental claims, always label the result status:

Status Meaning
Planned Not done yet. Can be described only as a plan.
Expected Hypothesis or target; not evidence.
Preliminary Early result that still needs validation.
Completed Finished and traceable to evidence.

Human Review Points

The agent can prepare drafts and evidence artifacts, but the human author must confirm:

  • final submission;
  • authorship;
  • ethics and compliance;
  • school or advisor requirements;
  • final claims and rebuttal commitments.

中文说明

这个 Skill 的核心价值是减少学术写作中的不可接受风险:虚构论文、虚构 SOTA、虚构实验结果、不受支持的强 claim,以及编造学校或导师要求。

不可妥协规则:

  1. 不虚构论文。
  2. 不虚构 SOTA。
  3. 不虚构结果。
  4. 不编造学校、学院、导师、模板、rubric、答辩或工作量要求。
  5. 优先使用 primary sources。
  6. 每个强 claim 都要映射到证据。
  7. 证据不足时降低或限定 claim。
  8. 只加载需要的模块。
  9. 外部 skill 只能低耦合辅助,不能覆盖完整性规则。
  10. 高风险决定需要人类作者复核。

写实验结果时必须标注状态:Planned、Expected、Preliminary 或 Completed。没有完成的实验不能写成已经证明。

Clone this wiki locally