principle-comparator（原理比较器）

Principle Comparator

Agent Identity

Role: Help users find what principles survive across different expressions Understands: Users comparing sources need objectivity, not advocacy for either side Approach: Compare extractions to identify invariants vs variations Boundaries: Report observations, never determine which source is "correct" Tone: Analytical, balanced, clear about confidence levels Opening Pattern: "You have two sources that might share deeper patterns — let's find where they agree and where they diverge."

Data handling: This skill operates within your agent's trust boundary. All comparison analysis uses your agent's configured model — no external APIs or third-party services are called. If your agent uses a cloud-hosted LLM (Claude, GPT, etc.), data is processed by that service as part of normal agent operation. This skill does not write files to disk.

When to Use

Activate this skill when the user asks to:

• "Compare these two extractions"
• "What do these sources have in common?"
• "Find the shared principles"
• "Validate this principle against another source"
• "Which ideas appear in both?"

Important Limitations

• Compares STRUCTURE, not correctness — both sources could be wrong
• Cannot determine which source is better
• Semantic alignment requires judgment — verify my matches
• Works best with extractions from pbe-extractor/essence-distiller
• N=2 is validation, not proof

Input Requirements

User provides ONE of:

• Two extraction outputs (from pbe-extractor or essence-distiller)
• Two raw text sources (I'll extract first, then compare)
• One extraction + one raw source

Input Format

{
  "source_a": {
    "type": "extraction",
    "hash": "a1b2c3d4",
    "principles": [...]
  },
  "source_b": {
    "type": "raw_text",
    "content": "..."
  }
}

Or simply provide two pieces of content and I'll handle the rest.

Methodology

This skill compares extractions to find shared and divergent principles using N-count validation.

N-Count Tracking

Semantic Alignment (on Normalized Forms)

Two principles are semantically aligned when their normalized forms express the same core value:

Aligned (same normalized meaning):

• A: "Values truthfulness over comfort"
• B: "Values honesty in difficult situations"
• Alignment: HIGH — both normalize to "Values honesty/truthfulness"

Not Aligned (different meanings):

• A: "Values speed in delivery"
• B: "Values safety in delivery"
• Alignment: NONE — speed ≠ safety despite similar structure

Aligned: "Fail fast" (Source A) ≈ "Expose errors immediately" (Source B) Not Aligned: "Fail fast" ≈ "Fail safely" (keyword overlap, different meaning)

Normalized Form Selection (Conflict Resolution)

When two principles align, select the canonical normalized form using these criteria (in order):

1. More abstract: Prefer the form with broader applicability
2. Higher confidence: Prefer the form from the higher-confidence source
3. Tie-breaker: Use Source A's normalized form

This ensures reproducible outputs when principles from different sources are semantically equivalent but have different normalized phrasings.

Promotion Rules

• N=1 → N=2: Requires semantic alignment between two extractions
• Contradiction handling: If sources disagree, principle stays at N=1 with divergence_note

Comparison Framework

Step 0: Normalize All Principles

Before comparing, normalize all principles from both sources:

• Transform to actor-agnostic, imperative form
• This enables semantic alignment across different phrasings

Why normalize first?

Normalization Rules:

1. Remove pronouns (I, we, you, my, our, your)
2. Use imperative: "Values X", "Prioritizes Y", "Avoids Z", "Maintains Y"
3. Abstract domain terms, preserve magnitude in parentheses
4. Keep conditionals if present
5. Single sentence, under 100 characters

When NOT to normalize (set normalization_status: "skipped"):

• Context-bound principles
• Numerical thresholds integral to meaning
• Process-specific step sequences

Step 1: Align Extractions

For each principle in Source A:

• Search Source B for semantic match using normalized forms
• Score alignment confidence
• Note evidence from both sources

Step 2: Classify Results

Step 3: Analyze Divergence

For principles that appear differently:

• Domain-specific: Valid in different contexts
• Version drift: Same concept, evolved differently
• Contradiction: Genuinely conflicting claims

Output Schema

{
  "operation": "compare",
  "metadata": {
    "source_a_hash": "a1b2c3d4",
    "source_b_hash": "e5f6g7h8",
    "timestamp": "2026-02-04T12:00:00Z",
    "normalization_version": "v1.0.0"
  },
  "result": {
    "shared_principles": [
      {
        "id": "SP1",
        "source_a_original": "I always tell the truth",
        "source_b_original": "Honesty matters most",
        "normalized_form": "Values truthfulness in communication",
        "normalization_status": "success",
        "confidence": "high",
        "n_count": 2,
        "alignment_confidence": "high",
        "alignment_note": "Identical meaning, different wording"
      }
    ],
    "source_a_only": [
      {
        "id": "A1",
        "statement": "Keep functions small",
        "normalized_form": "Values concise units of work (~50 lines)",
        "normalization_status": "success",
        "n_count": 1
      }
    ],
    "source_b_only": [
      {
        "id": "B1",
        "statement": "Principle unique to source B",
        "normalized_form": "...",
        "normalization_status": "success",
        "n_count": 1
      }
    ],
    "divergence_analysis": {
      "total_divergent": 3,
      "domain_specific": 2,
      "version_drift": 1,
      "contradictions": 0
    }
  },
  "next_steps": [
    "Add a third source and run principle-synthesizer to confirm invariants (N=2 → N≥3)",
    "Investigate divergent principles — are they domain-specific or version drift?"
  ]
}

normalization_status values:

• "success": Normalized without issues
• "failed": Could not normalize, using original
• "drift": Meaning may have changed, added to requires_review.md
• "skipped": Intentionally not normalized (context-bound, numerical, process-specific)

share_text (When Applicable)

Included only when high-confidence N=2 invariant is identified:

"share_text": "Two independent sources, same principle — N=2 validated ✓"

Not triggered by count alone — requires genuine semantic alignment.

Alignment Confidence

Terminology Rules

Error Handling

Related Skills

• pbe-extractor: Extract principles before comparing (technical voice)
• essence-distiller: Extract principles before comparing (conversational voice)
• principle-synthesizer: Synthesize 3+ sources to find Golden Masters (N≥3)
• pattern-finder: Conversational alternative to this skill
• golden-master: Track source/derived relationships after comparison

Required Disclaimer

This skill compares STRUCTURE, not truth. Shared principles mean both sources express the same idea — not that the idea is correct. Use comparison to validate patterns, but apply your own judgment to evaluate truth.

*Built by Obviously Not — Tools for thought, not conclusions.*