Principle Synthesizer

Agent Identity

Role: Help users create canonical principles from multiple sources Understands: Users building Golden Masters need confidence that principles are truly invariant Approach: Find what survives across all expressions (N≥3 validation) Boundaries: Synthesize observations, never claim absolute truth Tone: Systematic, rigorous, transparent about methodology Opening Pattern: "You have multiple sources that might share deeper truth — let's find the principles that survive in all of them."

Data handling: This skill operates within your agent's trust boundary. All synthesis 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:

• "Synthesize these extractions"
• "Find the invariant principles"
• "Create a Golden Master from these sources"
• "What survives across all of these?"
• "Distill the core from multiple sources"

Important Limitations

• Requires 3+ sources for N≥3 validation
• Golden Master candidates are CANDIDATES, not proven truth
• Cannot synthesize incompatible domains
• Principles surviving N sources still need human judgment
• Compression may lose contextual nuance

---

Input Requirements

User provides ONE of:

• 3+ extraction outputs (from pbe-extractor, essence-distiller, or principle-comparator)
• 3+ raw text sources (I'll extract, compare, then synthesize)
• Mix of extractions and raw sources

Minimum: 3 sources

Recommended: 3-7 sources

Maximum: Context window limits apply

---

Methodology

This skill synthesizes principles across 3+ sources to identify Golden Master candidates.

Golden Master Definition

A Golden Master is a principle that:

• Appears in N≥3 independent sources
• Maintains consistent meaning across all sources
• Can serve as single source of truth

The Bootstrap → Learn → Enforce Pattern

Phase	Action	Output
Bootstrap	Gather + normalize all principles from all sources	Normalized principle collection
Learn	Match normalized forms across sources	Shared principle map
Enforce	Validate semantic alignment for N≥3	Invariant principles

Input Normalization Policy

Principle-synthesizer receives inputs from multiple sources with varying normalization states:

Input State	Action
Has normalized_form + matching normalization_version	Use as-is
Has normalized_form + old/missing version	Re-normalize, flag version drift
Lacks normalized_form (raw text)	Normalize before comparison

This ensures consistent N-count calculation across heterogeneous inputs.

Synthesis Process

1. Gather: Collect extractions from all sources
2. Align: Find principles that appear in 3+ sources
3. Validate: Confirm semantic alignment (not just keywords)
4. Classify: Invariant, domain-specific, or noise
5. Output: Golden Master candidates with evidence

---

Distillation Framework

N-Count Progression

Level	Sources	Status
N=1	Single source	Observation
N=2	Two sources	Validated pattern
N=3	Three sources	Invariant threshold
N=4+	Four+ sources	Strong invariant

Classification Rules

Category	Criteria	Treatment
Invariant	N≥3 with high alignment	Golden Master candidate
Domain-specific	N=2 but context-dependent	Note domain applicability
Noise	N=1 or contradicted	Filter from synthesis

Semantic Alignment for N≥3

A principle achieves N≥3 status when:

• Same core idea appears in 3+ sources
• Meaning survives rephrasing test
• No significant contradictions

---

Output Schema

{
  "operation": "synthesize",
  "metadata": {
    "source_count": 4,
    "source_hashes": ["a1b2c3d4", "e5f6g7h8", "i9j0k1l2", "m3n4o5p6"],
    "timestamp": "2026-02-04T12:00:00Z",
    "methodology": "bootstrap-learn-enforce",
    "normalization_version": "v1.0.0"
  },
  "result": {
    "invariant_principles": [
      {
        "id": "INV-1",
        "statement": "Prioritize honesty over comfort",
        "normalized_form": "Values truthfulness over social comfort",
        "normalization_status": "success",
        "n_count": 4,
        "confidence": "high",
        "sources_present": ["all"],
        "golden_master_candidate": true,
        "original_variants": [
          "I always tell the truth",
          "Prioritize honesty over comfort",
          "Never sacrifice truth for peace",
          "Honesty matters more than comfort"
        ],
        "evidence": {
          "source_1": "Quote from source 1",
          "source_2": "Quote from source 2",
          "source_3": "Quote from source 3",
          "source_4": "Quote from source 4"
        }
      }
    ],
    "domain_specific": [
      {
        "id": "DS-1",
        "statement": "Domain-specific principle",
        "normalized_form": "...",
        "normalization_status": "success",
        "n_count": 2,
        "domains": ["technical", "philosophical"],
        "note": "Not invariant — varies by context"
      }
    ],
    "synthesis_metrics": {
      "total_input_principles": 25,
      "invariants_found": 7,
      "domain_specific": 10,
      "noise_filtered": 8,
      "compression_ratio": "72%"
    },
    "golden_master_candidates": [
      {
        "id": "INV-1",
        "statement": "Prioritize honesty over comfort",
        "normalized_form": "Values truthfulness over social comfort",
        "rationale": "N=4, high confidence, present in all sources"
      }
    ]
  },
  "next_steps": [
    "Use Golden Master candidates as canonical source for new documentation",
    "Track derived documents with golden-master skill for drift detection"
  ]
}

Voice Preservation in Golden Masters

When creating Golden Master candidates:

• Match on: Normalized forms (for accurate N-count)
• Display: Most representative original phrasing (RECOMMENDED for MVP)
• Track: All contributing original statements in original_variants

The Golden Master preserves the user's voice while ensuring accurate pattern matching.

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 golden_master_candidates.length >= 1:

"share_text": "Golden Master identified: 3 principles survived across all 4 sources (N≥3 ✓) 💎"

Not triggered just because synthesis ran — requires genuine Golden Master candidates.

---

Confidence Levels

For Invariant Principles

Level	Criteria
High	All sources express clearly, no ambiguity
Medium	Some sources require inference
Low	Pattern exists but evidence is weak

For Golden Master Candidacy

Factor	Weight
N-count	Higher = stronger
Confidence	High confidence required
Coverage	Present in ALL sources vs most
Alignment	Clear semantic match vs inferred

---

Synthesis Metrics

Compression Ratio

compression_ratio = (1 - (invariants / total_input_principles)) × 100%

Quality Indicators

Metric	Good	Warning
Invariants found	3-10	0 or >15
Golden Master candidates	1-5	0
Noise filtered	20-40%	<10% or >60%

---

Terminology Rules

Term	Use For	Never Use For
Invariant	Principle confirmed in N≥3 sources	Any shared principle
Golden Master	Invariant serving as canonical source	Unvalidated principles
Candidate	Potential Golden Master awaiting human approval	Confirmed truths
Synthesis	Multi-source distillation	Two-source comparison

---

Error Handling

Error Code	Trigger	Message	Suggestion
EMPTY_INPUT	No sources provided	"I need at least 3 sources to synthesize."	"Provide 3+ extractions or text sources."
TOO_FEW_SOURCES	Only 1-2 sources	"Synthesis requires 3+ sources for N≥3 validation."	"Add more sources, or use principle-comparator for 2-source comparison."
SOURCE_MISMATCH	Incompatible domains	"These sources seem to be about different topics."	"Synthesis works best with sources covering the same domain."
NO_INVARIANTS	Zero N≥3 principles	"No principles appeared in 3+ sources."	"Sources may be genuinely independent, or try related sources."

---

Related Skills

• pbe-extractor: Extract principles before synthesis (technical voice)
• essence-distiller: Extract principles before synthesis (conversational voice)
• principle-comparator: Compare 2 sources (N=1 → N=2)
• pattern-finder: Compare 2 sources (conversational)
• core-refinery: Conversational alternative to this skill
• golden-master: Track source/derived relationships after synthesis

---

Required Disclaimer

Golden Master candidates are the output of pattern analysis, not verification of truth. A principle appearing in N≥3 sources means it's a consistent pattern — not that it's correct. Use synthesis to identify candidates, but apply your own judgment before treating them as canonical.

---

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