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#

PhaseActionOutputBootstrapGather + normalize all principles from all sourcesNormalized principle collectionLearnMatch normalized forms across sourcesShared principle mapEnforceValidate semantic alignment for N≥3Invariant principles

Input Normalization Policy#

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

Input StateActionHas normalized_form + matching normalization_versionUse as-isHas normalized_form + old/missing versionRe-normalize, flag version driftLacks 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#

LevelSourcesStatusN=1Single sourceObservationN=2Two sourcesValidated patternN=3Three sourcesInvariant thresholdN=4+Four+ sourcesStrong invariant

Classification Rules#

CategoryCriteriaTreatmentInvariantN≥3 with high alignmentGolden Master candidateDomain-specificN=2 but context-dependentNote domain applicabilityNoiseN=1 or contradictedFilter 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#

json

{
  "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:

json

"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#

LevelCriteriaHighAll sources express clearly, no ambiguityMediumSome sources require inferenceLowPattern exists but evidence is weak

For Golden Master Candidacy#

FactorWeightN-countHigher = strongerConfidenceHigh confidence requiredCoveragePresent in ALL sources vs mostAlignmentClear semantic match vs inferred

---

Synthesis Metrics#

Compression Ratio#

text

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

Quality Indicators#

MetricGoodWarningInvariants found3-100 or >15Golden Master candidates1-50Noise filtered20-40%<10% or >60%

---

Terminology Rules#

TermUse ForNever Use ForInvariantPrinciple confirmed in N≥3 sourcesAny shared principleGolden MasterInvariant serving as canonical sourceUnvalidated principlesCandidatePotential Golden Master awaiting human approvalConfirmed truthsSynthesisMulti-source distillationTwo-source comparison

---

Error Handling#

Error CodeTriggerMessageSuggestionEMPTY_INPUTNo sources provided"I need at least 3 sources to synthesize.""Provide 3+ extractions or text sources."TOO_FEW_SOURCESOnly 1-2 sources"Synthesis requires 3+ sources for N≥3 validation.""Add more sources, or use principle-comparator for 2-source comparison."SOURCE_MISMATCHIncompatible domains"These sources seem to be about different topics.""Synthesis works best with sources covering the same domain."NO_INVARIANTSZero 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.