feat: add emulator agent for mind clone activation

core/protocols/dna-mental-extraction.md

@@ -0,0 +1,396 @@
+# DNA Mental Extraction Protocol
+
+> **Version:** 1.0.0
+> **Schema:** `core/schemas/dna-mental-8-layer.schema.json`
+> **Glossary:** `core/glossary/dna-mental-layers.md`
+> **Scope:** Extraction of all 8 DNA layers from source materials
+
+---
+
+## Purpose
+
+This protocol governs how the 8 layers of DNA Mental are extracted from raw source materials (transcriptions, books, courses, podcasts). It defines the order of extraction, quality gates, and validation checkpoints for each layer.
+
+---
+
+## Extraction Phases
+
+```
+  PHASE 1: EXPLICIT EXTRACTION (Layers 1-5)
+  ==========================================
+  Single-pass extraction from source material.
+  Each batch of source material yields L1-L5 items directly.
+
+      Source Material
+           │
+           ▼
+  ┌────────────────────┐
+  │  L5 Methodologies  │◄── "Step 1... Step 2..." patterns
+  ├────────────────────┤
+  │  L4 Frameworks     │◄── Named systems, multi-component models
+  ├────────────────────┤
+  │  L3 Heuristics     │◄── Rules, thresholds, "always/never"
+  ├────────────────────┤
+  │  L2 Mental Models  │◄── "Think of it like...", analogies
+  ├────────────────────┤
+  │  L1 Philosophies   │◄── "I believe...", "The truth is..."
+  └────────────────────┘
+
+  PHASE 2: DEEP INFERENCE (Layers 6-8)
+  ==========================================
+  Cross-source analysis AFTER all sources for a person are processed.
+  Requires the complete L1-L5 corpus as input.
+
+      Complete L1-L5 Corpus
+           │
+           ▼
+  ┌────────────────────┐
+  │  L6 Values         │◄── Trade-off analysis across all sources
+  ├────────────────────┤
+  │  L7 Obsessions     │◄── Frequency analysis across all sources
+  ├────────────────────┤
+  │  L8 Paradoxes      │◄── Contradiction detection + resolution
+  └────────────────────┘
+           │
+           ▼
+      HUMAN VALIDATION (L8 only)
+```
+
+---
+
+## Phase 1: Explicit Extraction (L1-L5)
+
+### When to Execute
+
+- During Pipeline Phase 4 (batch processing)
+- For each batch of source material
+
+### Extraction Order
+
+Process each batch in this order (most concrete first):
+
+1. **L5 Methodologies** -- Scan for step-by-step instructions
+2. **L4 Frameworks** -- Scan for named models and structured systems
+3. **L3 Heuristics** -- Scan for rules, thresholds, decision shortcuts
+4. **L2 Mental Models** -- Scan for analogies, lenses, thinking tools
+5. **L1 Philosophies** -- Scan for beliefs, convictions, worldview statements
+
+### Signal Words by Layer
+
+| Layer | Signal Words / Patterns |
+|-------|------------------------|
+| L5 | "Step 1", "First you", "The process is", "How to", "Do X then Y" |
+| L4 | Named acronyms (CLOSER, SPIN), "The X framework", "My system for", components listed |
+| L3 | Numbers + context ("8-12%", "3x minimum"), "Always", "Never", "Rule of thumb", "If X then Y" |
+| L2 | "Think of it like", "It's similar to", "The way I see it", "Imagine", analogies |
+| L1 | "I believe", "The truth is", "What people don't understand", "At the end of the day", "The fundamental thing" |
+
+### Quality Gates (L1-L5)
+
+For each extracted item, verify:
+
+```
+[ ] Item has a valid ID in the correct pattern
+[ ] Item has at least 1 source_reference with chunk_id
+[ ] Item's source quote is a direct extraction (not paraphrased)
+[ ] Item is classified in the correct layer (not misplaced)
+[ ] Item does not duplicate an existing item (check by semantic similarity)
+[ ] Confidence score reflects extraction certainty
+```
+
+### Deduplication Rules
+
+When a new item seems to duplicate an existing one:
+
+| Scenario | Action |
+|----------|--------|
+| Exact same concept, same person | Increment `reinforcement_count`, add new source |
+| Same concept, different person | Create separate items (different ID prefix), note cross-reference |
+| Similar concept, nuanced difference | Create separate items with clear distinction in description |
+| Contradictory concept, same person | Flag for L8 Paradox analysis in Phase 2 |
+
+---
+
+## Phase 2: Deep Inference (L6-L8)
+
+### When to Execute
+
+- ONLY after ALL source material for a person has been processed through Phase 1
+- During Pipeline Phase 5 (agent generation), sub-phase 5.1 (Foundation)
+- Requires the complete L1-L5 corpus for that person as input
+
+### Prerequisites
+
+```
+[ ] All batches for person XX processed through Phase 1
+[ ] Complete L1-L5 corpus available (DNA.yaml populated)
+[ ] Source count >= 3 (minimum for triangulation)
+[ ] No pending batch processing for this person
+```
+
+---
+
+### L6 Extraction: Values Hierarchy
+
+**Method:** Trade-off analysis across all documented decisions and behaviors.
+
+**Step-by-step:**
+
+1. **Collect decision points** -- Scan all L1-L5 items and source material for moments where the person made explicit choices between competing priorities.
+
+2. **Build trade-off pairs** -- For each decision, document:
+   - What they CHOSE (reveals the higher value)
+   - What they SACRIFICED (reveals the lower value)
+   - The context of the decision
+
+3. **Cluster into values** -- Group trade-off pairs by the underlying value being served:
+   - Speed, Control, Freedom, Impact, Mastery, Wealth, Security, etc.
+
+4. **Rank by consistency** -- The value that wins most consistently in trade-offs ranks highest.
+
+5. **Cross-validate with L1** -- Compare revealed values against declared philosophies (L1). Note alignments and contradictions.
+
+6. **Require triangulation** -- Each value must have evidence from 3+ independent sources.
+
+**Evidence collection template:**
+
+```yaml
+value: "Speed"
+trade_off_evidence:
+  - chose: "[specific thing chosen]"
+    over: "[specific thing sacrificed]"
+    context: "[when/where this happened]"
+    source:
+      source_id: "AH003"
+      chunk_id: "AH003-045"
+      quote: "[exact words]"
+```
+
+**Quality gate:**
+
+```
+[ ] Each value has >= 3 independent trade-off examples
+[ ] Trade-offs come from >= 3 different sources
+[ ] Ranking is consistent (no value wins AND loses equally)
+[ ] declared_vs_revealed field is set accurately
+[ ] No values based solely on self-declarations
+```
+
+---
+
+### L7 Extraction: Core Obsessions
+
+**Method:** Frequency analysis + cross-layer manifestation mapping.
+
+**Step-by-step:**
+
+1. **Frequency scan** -- Analyze ALL processed source material. Identify themes/drivers that appear in >= 50% of sources.
+
+2. **Distinguish topics from obsessions:**
+   - TOPIC: "Sales" -- this is a domain, not an obsession
+   - OBSESSION: "Leverage" -- this is the WHY behind sales frameworks that minimize input and maximize output
+
+3. **Map manifestations** -- For each candidate obsession, find where it appears across layers:
+   - Does it appear in L1 Philosophies?
+   - Does it shape L2 Mental Models?
+   - Does it create L3 Heuristics?
+   - Does it structure L4 Frameworks?
+   - Does it drive L5 Methodology design?
+   - Does it explain L6 Value trade-offs?
+
+4. **Require cross-layer presence** -- Must manifest in >= 3 different layers.
+
+5. **Limit to 2-5 items** -- If analysis yields more than 5 candidates, go deeper. More than 5 means the analysis is too surface-level.
+
+6. **Document origin hypothesis** -- Optionally note biographical/experiential roots.
+
+**Quality gate:**
+
+```
+[ ] Each obsession appears in >= 50% of source material
+[ ] Each obsession manifests in >= 3 different layers (L1-L6)
+[ ] Total obsessions: 2-5 (not more)
+[ ] Obsessions are DRIVERS, not topics or domains
+[ ] Each has >= 3 independent source references
+[ ] Frequency is calculated, not estimated
+```
+
+---
+
+### L8 Extraction: Productive Paradoxes
+
+**Method:** Contradiction detection + contextual resolution.
+
+**Step-by-step:**
+
+1. **Scan for contradictions** -- Review all L1-L5 items for the same person. Flag pairs where:
+   - Item A says "do X" and Item B says "do NOT-X"
+   - Heuristic A has threshold > N and Heuristic B implies threshold < N
+   - Philosophy A and Philosophy B seem incompatible
+
+2. **Filter out non-paradoxes:**
+   - Evolution over time: "Used to think X, now thinks Y" -- NOT a paradox
+   - Audience difference: "Says X to beginners, Y to experts" -- may be pedagogical
+   - Imprecision: "Said 8% once and 12% another time" -- NOT a paradox
+   - Context already encoded: "Do X in situation A, do Y in situation B" -- already resolved
+
+3. **Confirm genuine paradoxes** -- For remaining candidates:
+   - Verify BOTH sides have >= 2 independent sources
+   - Verify the person genuinely holds BOTH positions simultaneously (not sequentially)
+   - Verify the context difference is real (not just different wording)
+
+4. **Write the synthesis** -- For each confirmed paradox, articulate:
+   - WHEN the thesis applies (thesis_context)
+   - WHEN the antithesis applies (antithesis_context)
+   - HOW both coexist (synthesis)
+
+5. **Set validation_status to `pending_human_review`** -- NEVER set to `human_confirmed` automatically.
+
+6. **Optional: auto_high_confidence** -- If BOTH sides have 3+ sources AND confidence > 0.95, may flag as `auto_high_confidence`. This allows provisional use with a disclaimer but does NOT replace human validation.
+
+**Quality gate:**
+
+```
+[ ] Each paradox has >= 2 sources per side (4 total minimum)
+[ ] Non-paradoxes filtered out (evolution, imprecision, audience)
+[ ] Synthesis explains contextual coexistence clearly
+[ ] thesis_context and antithesis_context are distinct and specific
+[ ] validation_status is "pending_human_review" (NOT "human_confirmed")
+[ ] Related items in other layers are cross-referenced
+```
+
+---
+
+## Human Validation Protocol (L8)
+
+### Who Validates
+
+The human stakeholder who deeply understands the expert being modeled. This is NOT a casual review -- it requires someone who can distinguish genuine paradoxes from extraction errors.
+
+### Validation Checklist
+
+For each L8 item in `pending_human_review`:
+
+```
+[ ] Is this a GENUINE contradiction the expert holds?
+    → YES: Set to "human_confirmed"
+    → NO: Set to "human_rejected" with reason
+
+[ ] Is the synthesis accurate?
+    → If not, revise synthesis before confirming
+
+[ ] Are the contexts correct?
+    → Thesis context: Is this really when side A applies?
+    → Antithesis context: Is this really when side B applies?
+
+[ ] Would you be comfortable if the mind clone used this paradox?
+    → If hesitant, keep in "pending_human_review"
+```
+
+### After Validation
+
+- `human_confirmed` items are published to agents and used in reasoning
+- `human_rejected` items are logged with rejection reason and archived
+- `pending_human_review` items are NOT used in agent generation
+
+---
+
+## Integration with Existing Pipeline
+
+### Phase 4 (Batch Processing) Changes
+
+Each batch now extracts L1-L5 items. The batch log template adds:
+
+```
+┌─ DNA EXTRACTION ────────────────────────────────────────────────────────────┐
+│  L1 Philosophies:   +N new, M reinforced                                    │
+│  L2 Mental Models:  +N new, M reinforced                                    │
+│  L3 Heuristics:     +N new (Q quantitative, R qualitative), M reinforced    │
+│  L4 Frameworks:     +N new, M reinforced                                    │
+│  L5 Methodologies:  +N new, M reinforced                                    │
+│                                                                             │
+│  L6-L8: Deferred to Phase 5 (requires complete corpus)                      │
+└─────────────────────────────────────────────────────────────────────────────┘
+```
+
+### Phase 5 (Agent Generation) Changes
+
+Sub-phase 5.1 (Foundation) now includes L6-L8 extraction:
+
+```
+5.1 FOUNDATION (per source):
+    5.1.1  Load all batches for source
+    5.1.2  Consolidate L1-L5 DNA
+    5.1.3  Extract L6 Values Hierarchy    ← NEW
+    5.1.4  Extract L7 Core Obsessions     ← NEW
+    5.1.5  Extract L8 Productive Paradoxes ← NEW
+    5.1.6  Queue L8 for human validation   ← NEW
+    5.1.7  Create SOURCE-XX.md
+```
+
+### DNA-CONFIG.yaml Changes
+
+The `dna_sources` section in DNA-CONFIG.yaml should reference which layers have been extracted:
+
+```yaml
+dna_sources:
+  primario:
+    - pessoa: "alex-hormozi"
+      layers_extracted: ["L1", "L2", "L3", "L4", "L5", "L6", "L7", "L8"]
+      l8_validation_status: "pending_human_review"  # or "human_confirmed"
+```
+
+---
+
+## Reasoning Cascade Update
+
+The agent reasoning cascade (AGENT-COGNITION-PROTOCOL) is extended for L6-L8:
+
+```
+EXISTING CASCADE (L1-L5):
+  METHODOLOGY → FRAMEWORK → HEURISTICS → MENTAL MODEL → PHILOSOPHY
+
+EXTENDED CASCADE (L1-L8):
+  METHODOLOGY (L5)
+      → FRAMEWORK (L4)
+      → HEURISTICS (L3)
+      → MENTAL MODEL (L2)
+      → PHILOSOPHY (L1)
+      → VALUES CHECK (L6): "Does this align with revealed values?"
+      → OBSESSION CHECK (L7): "Is a core obsession influencing this?"
+      → PARADOX CHECK (L8): "Is there a productive paradox here?"
+
+L6 CHECK: After forming a recommendation, verify it aligns with the
+          expert's revealed values hierarchy. If it conflicts with a
+          top-ranked value, flag and explain.
+
+L7 CHECK: After forming a recommendation, check if a core obsession
+          is pulling the reasoning in a specific direction. If so,
+          make it transparent: "Note: This recommendation reflects
+          [PERSON]'s core obsession with [X]."
+
+L8 CHECK: If the situation maps to a known productive paradox,
+          present BOTH sides with their contexts instead of picking
+          one side. This is what makes the clone feel human.
+```
+
+---
+
+## Error Handling
+
+| Error | Resolution |
+|-------|------------|
+| L6 value has < 3 sources | Do NOT publish. Mark as `draft` and collect more evidence |
+| L7 obsession frequency < 50% | Demote to "recurring theme" (not obsession). Do not include in L7 |
+| L7 yields > 5 obsessions | Re-analyze at deeper level. Merge overlapping obsessions |
+| L8 paradox is actually evolution | Reclassify as temporal change. Remove from L8 |
+| L8 paradox is imprecision | Reclassify as data quality issue. Remove from L8 |
+| L8 human rejects paradox | Archive with rejection reason. Do not re-submit same paradox |
+
+---
+
+## Changelog
+
+| Version | Date | Change |
+|---------|------|--------|
+| 1.0.0 | 2026-02-28 | Initial extraction protocol for 8-layer DNA Mental |