360 Secure SkillHub
360 安全技能中心

Red Team

在把数值结论从“初步观察”升级为“可报告结果”前，按文献基线、可观测审计、断言分类、可证伪段落、同行质疑五道门禁逐条书面作答；任一门禁缺失即保持暂定，并提供日志模板记录判定。

space-cadet
space-cadet
热度 29.49
最近更新 2026-07-23 14:20:19
分类：知识依赖：纯提示词应用场景：研究 / 报告撰写

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add space-cadet/red-team
Metadata
name
red-team
description
Red-team protocol for validating numerical claims before promoting them from tentative to real. 5 gates: literature baseline, observable audit, claim taxonomy,

Red-Team Protocol

Purpose

Before promoting any numerical claim from "tentative observation" to "finding worth reporting, committing, or dashboarding," run this protocol. No exceptions. The red team is a skeptical colleague who will tear your claim apart if you let them.

Trigger

Any claim involving:

  • Numerical results from computation or simulation
  • Comparisons to literature values
  • Physical predictions or bounds
  • Trend identification ("increases with", "decreases with")
  • Statistical summaries ("on average", "typical value")

The 5 Gates

Answer each gate explicitly, in writing, before calling the claim a "result."

Gate 1: Literature Baseline

Question: What does the literature say? What's the accepted value or range?

Required:

  • Cite at least one source (paper, book, arXiv, review)
  • State the accepted value with uncertainty if available
  • Note if your result contradicts, agrees with, or extends known results

If no literature exists: State explicitly that this is a prediction with no known comparison.

If you cannot find a source: The claim stays tentative until you do.

Gate 2: Observable Audit

Question: What did I actually compute? Method, parameters, sample size, error bars.

Required:

  • Exact method name (algorithm, solver, approximation scheme)
  • All parameters used (no defaults assumed without checking)
  • Sample size / number of data points / convergence level
  • Statistical uncertainty or error estimate
  • If the computation was interrupted or resumed, say so

No hand-waving. "The usual parameters" or "standard settings" are not acceptable. Write the numbers down.

Gate 3: Claim Taxonomy

Question: What kind of claim is this?

Classify exactly one:

Type Example Required Evidence
Point estimate "The ground state energy is 0.34" Convergence plot, error bars
Trend "Energy increases with g" Multiple data points, monotonicity check
Existence proof "A phase transition exists" Finite-size scaling, multiple L values
Bound "The error is less than 0.01" Upper bound derivation, checked numerically
Null result "No phase transition found" Scan of parameter space, finite-size analysis
Comparison "Our result agrees with X" Quantitative difference, error propagation

If you cannot classify the claim: It is not a result. It is an observation. Go back to work.

Gate 4: Falsification Paragraph

Question: What would prove this wrong? What could I check that I'm not checking?

Required: Write 2-3 sentences describing:

  • A plausible error source that would invalidate the claim
  • A check you haven't done but could do
  • A limit or caveat that makes the claim provisional

Example: "If the finite-size scaling is wrong because the correlation length exceeds L=12, the trend in g_c(L) would be an artifact. I should check ξ vs L before claiming convergence."

If you cannot identify a falsification path: You are not thinking hard enough. Try again.

Gate 5: Peer-Check Gate

Question: If I had to explain this to a skeptical colleague, what's the weakest point?

Required: State the weakest point in 1-2 sentences. Then ask:

  • Would this pass a peer review? (yes/no/close)
  • What would Reviewer 2 complain about?

If the weakest point is the entire claim: The claim is not ready. Back to Gate 1.

Enforcement

  1. Write the answers out. Do not run the protocol mentally. Write it in the same file, message, or note where you are reporting the result.
  2. If any gate is missing or answered with "I don't know" or "I assume" — the claim stays TENTATIVE. No exceptions.
  3. No promotion without all 5 gates. A result without a red-team pass is a rumor, not science.
  4. When reporting to the user: Use language that reflects the gate status:
    • All gates passed: "The result is..." (confident)
    • Some gates partial: "The tentative result is..." (cautious)
    • No gates run: "I observed..." (descriptive only)

Red-Team Log

When a claim is promoted from tentative to confirmed, append a brief record to the red-team log (can be in the same memory file or a separate note):

[DATE] Claim: [one-line summary]
- Gate 1: [status] [source]
- Gate 2: [status] [method, params, N]
- Gate 3: [type]
- Gate 4: [falsification risk]
- Gate 5: [weakest point] [review-readiness]
- Verdict: CONFIRMED / TENTATIVE / REJECTED

Failure Modes

These are the ways this protocol has failed in the past. Watch for them:

  • Gate 1 skipped: "I didn't look it up because I know the answer." → You don't. Look it up.
  • Gate 2 hand-waved: "The usual parameters." → Write them down.
  • Gate 3 misclassified: Calling a trend a point estimate because you like the middle value. → Classify honestly.
  • Gate 4 weak: "Maybe I made a mistake." → Be specific. What mistake? How would you detect it?
  • Gate 5 optimistic: "I'm sure this is right." → Find the weakest point. It's there.

Related

  • self-improving-agent skill: Log red-team failures as learnings
  • spike skill: Run throwaway tests to validate claims before red-teaming