Analogy Safeguard

Changelog / 变更日志

AnalogySafeguard

历史类比安全检查

Validate that historical and classical analogies in debate turns comply with structural requirements and proportion limits. 验证辩论回合中的历史和经典类比是否符合结构要求和比例限制。

When to Use / 何时使用

• During DebateTurn construction as self-check / DebateTurn 构建时作为自检
• During JudgeAudit to validate analogy usage / JudgeAudit 验证类比使用时
• When reviewing any argument that references historical precedent / 审查任何引用历史先例的论点时

Input / 输入 (v3 update)

• content: The debate turn content to validate
• mode: "strict" | "advisory" (default: "strict")
  • strict: Used for analogies in arguments[] — full rules apply (≥2 similarities, ≥1 difference, <15% content share)
  • advisory: Used for historical_wisdom section — relaxed rules (requires key_differences and applicability_caveat, but no content share limit)

Output / 输出

• Per analogy: pass | fail with reasons
• Overall content share assessment (strict mode only)
• Recommendations for failed analogies

Validation Rules / 验证规则

For mode = "strict" (arguments[]):

Current behavior unchanged:

Rule 1: Structural Requirements / 结构要求

Each historical or classical analogy MUST include:

At least 2 similarities / 至少 2 个相似点:

• Structural parallels between the historical case and current topic
• Not superficial similarities (e.g., "both involve money" is too vague)
• Must be causally relevant similarities

At least 1 key structural difference / 至少 1 个关键结构差异:

• An honest acknowledgment of where the analogy breaks down
• Must be substantive, not trivial
• This prevents over-reliance on historical parallels

If an analogy lacks either → fail, mark as "heuristic only"

Rule 2: Content Share / 内容占比

Total analogy and historical-reference content should be under ~15% of the debate turn's total content.

This is assessed by LLM semantic judgment, not mechanical word count:

• Estimate what proportion of the argument's substance relies on historical parallels
• A brief reference to a historical event is fine
• Building an entire argument primarily on historical analogy is too much

If content share exceeds ~15% → flag (not auto-fail; Judge decides)

Rule 3: Analogy Classification / 类比分类

Analogies that pass structural validation:

• Valid supporting evidence in the reasoning track
• Can be cited as part of the mechanism explanation

Analogies that fail structural validation:

• Marked as "heuristic only" by Judge
• Cannot serve as primary evidence for a claim
• The debater should be told to strengthen with non-analogy evidence

For mode = "advisory" (historical_wisdom):

1. Verify key_differences field is substantive (not a token acknowledgment)
2. Verify applicability_caveat field is present and meaningful
3. NO content share limit — historical references can be explored in depth
4. Grade quality: "strong_parallel" | "moderate_parallel" | "weak_parallel"
5. If weak_parallel: suggest in output that the reference may not add value

Workflow / 工作流

1. Identify: Use LLM to find all analogy sections in the debate turn
2. Determine mode: Check if the analogy is in arguments[] (strict) or historical_wisdom (advisory)
3. Evaluate: For each analogy, assess according to the appropriate mode
4. Estimate: Assess overall content share semantically (strict mode only)
5. Report: Return per-analogy pass/fail and overall assessment

Examples of Good vs. Bad Analogies / 好坏类比示例

Good / 好: "The 2008 financial crisis (similarity 1: credit overextension, similarity 2: regulatory lag) suggests similar risks today, though the difference is that current regulatory frameworks are stronger post-Dodd-Frank."

Bad / 坏: "This is just like the 1929 crash" (no similarities articulated, no differences acknowledged)

Bad / 坏: "History always repeats itself" (not a structured analogy at all)