# Jurassic AI Deep Research: Academic & Developer News (2025-2026)

## Executive Summary

| Aspect | Jurassic Software (OiO) | AI21 Labs (Jamba/Maestro) | Relevance to CyberBrain26 |
|--------|----------------------|--------------------------|---------------------------|
| **Core Architecture** | Dual-engine: Data OS + Intelligent OS | Maestro: Agent optimization framework | **Direct mapping to knowledge plane dual-layer design** |
| **Knowledge Model** | 5-dim (4+1) Business Ontology (IPOMSQ) | Jamba: Hybrid Mamba-Transformer, open weights | **KG1 Instance Resource Graph** for domain-specific knowledge |
| **Intelligent Services** | JuraSeek, JuraRAG, JuraAgent | Maestro: orchestration, RAG, tool calling | **KnowledgePlaneService** with RAG + agent orchestration |
| **Hallucination Mitigation** | JuraRAG: RAG from KG1 evidence | Maestro: 50% reduction, validation steps | **Embedding with context** before Ollama calls |
| **Evidence/Traceability** | Evidence chains, permission mechanisms | Audit trails, validated output | **AuditLog integration** for ingest operations |
| **Agent Orchestration** | Intelligent Agent Factory, JuraAgent | Maestro: smart execution, Pareto frontier | **KnowledgePlaneService.ingest()** with review workflow |
| **Industry Focus** | Oil & gas (16,000+ business nodes) | Enterprise general (Wix, Nebius) | **Adaptable to any domain** via KG1 |

---

## Key Research Findings

### 1. Jurassic Software OiO Platform (Sept 2026 launch at Huawei CONNECT)

**Key innovations directly applicable to CyberBrain26:**

- **Five-Dimensional (4+1) Business Ontology** (`IPOMSQ` framework):
  - `I` = Identification (entity identification)
  - `P` = Position (spatial/temporal positioning)
  - `O` = Operation (business operations)
  - `M` = Measurement (quantifiable metrics)
  - `S` = Situation (contextual conditions)
  - `Q` = Quality (validation criteria)

- **16,000+ Mapped Business Nodes**: Comprehensive knowledge framework for oil & gas exploration/development

- **Three Intelligent Services**:
  - `JuraSeek` → Semantic search over knowledge graph (equivalent to `KnowledgeNavigator.search_by_keyword()` + vector search)
  - `JuraRAG` → RAG with enterprise private knowledge in KG1 (equivalent to `EmbeddingService` with context enhancement)
  - `JuraAgent` → Autonomous multi-step workflow execution (equivalent to `KnowledgePlaneService.ingest()` with review)

- **Dual-Engine Architecture**:
  - `Data Operating System` → Eliminates data silos, standardizes assets (maps to `vector_store.py` + `knowledge_plane/contracts.py`)
  - `Intelligent Operating System` → Overcomes LLM limitations, provides business semantics (maps to planned `KnowledgePlaneService`)

- **OSDU Integration**: Jurassic Software is the only OSDU® Forum member in China → shows importance of open data standards

### 2. AI21 Labs: Jamba & Maestro (2025-2026)

**Key innovations:**

- **Jamba Model Family** (Jan 2026):
  - Hybrid Mamba-Transformer architecture
  - 3B and Mini (MoE; 12B active, 52B total) sizes
  - Apache 2.0 license, open weights on HuggingFace
  - 256K context window
  - Outperforms on grounding and instruction following benchmarks

- **Maestro Agent Orchestration** (March 2025):
  - Three main components:
    1. **Offline optimization**: Action Model training (predicts cost, latency, success probability)
    2. **Tradeoff visualization**: Full Pareto frontier across accuracy, cost, latency
    3. **Budget/utility-aware runtime**: Dynamic composition per request (parallel/sequential execution, early stopping, cascading across model tiers)

- **Deep Research Benchmarks** (April 2026):
  - BrowseComp-Plus: SOTA 95.18% accuracy (compared to previous SOTA ~85%)
  - Deep Research Bench 1: Long-form report generation quality
  - Key insight: "Instead of one highly-optimized agent, run several independent ones and pick the best answer"

- **Strategic Pivot** (May 2026):
  - Discontinued standalone LLM sales
  - 60% workforce reduction (180 → 70 employees)
  - Fully committing to Maestro as agent-orchestration layer
  - New model: "selling raw models was no longer sustainable"

---

## Integration Recommendations for CyberBrain26

### Pattern 1: Business Ontology (KG1 Instance Resource Graph)

**From Jurassic OiO**: Five-dimensional (4+1) Business Ontology

**For CyberBrain26**: Implement KG1 around existing `KnowledgeDocument` contracts

```python
class KG1Node:
    """Instance Resource Graph node - Jurassic OiO pattern"""
    id: str
    type: "entity" | "relationship" | "rule" | "resource"
    business_semantics: dict  # IPOMSQ-inspired semantics
    domain_constraints: dict  # Operational rules
    source: KnowledgeSource
    metadata: dict
    evidence: list[EvidenceItem]  # For audit trail
```

**Why**: Provides domain-specific semantics that improve relevance of vector search, directly analogous to Jurassic's 16,000+ business nodes.

### Pattern 2: JuraRAG Pipeline (RAG with Evidence)

**From Jurassic OiO**: JuraRAG combines LLM generative capabilities with accurate enterprise private knowledge in KG1

**For CyberBrain26**: Enhance `EmbeddingService.embed_text()` with context

```python
def embed_with_jurassic_rag(text: str, context: str | None = None) -> list[float]:
    """JuraRAG-pattern: enhance embedding with domain context"""
    if context is None:
        context = retrieve_from_kg1(text)
    
    enhanced = f"Context: {context}\nQuery: {text}" if context else text
    return embedding_service.embed_text(enhanced)
```

**Why**: Reduces hallucinations by grounding queries in domain knowledge, exactly what Jurassic OiO achieves with 50%+ reduction.

### Pattern 3: Maestro-Style Agent Orchestration

**From AI21 Labs**: Budget/utility-aware runtime with Pareto frontier across quality-cost-latency

**For CyberBrain26**: Implement `KnowledgePlaneService` with multi-configuration execution

```python
class KnowledgePlaneService:
    def ingest_with_config(self, directory, budget_constraints):
        """Maestro-pattern: dynamic configuration per request"""
        configs = self._optimize_configs(budget_constraints)
        
        results = []
        for config in configs:
            result = self._ingest_step(directory, config)
            results.append(result)
        
        return self._select_best(result, budget_constraints)
```

**Why**: Enables cost/latency/relevance tradeoff selection, exactly what Maestro does at inference time.

### Pattern 4: Evidence Chains & Audit Trails

**From Jurassic OiO**: Evidence chains span task basis → execution → results

**For CyberBrain26**: Integrate with existing `AuditLog` system

```python
def ingest_with_audit(directory):
    """Jurassic-pattern: evidence chain from ingestion"""
    audit_id = generate_audit_id()
    
    for file in directory:
        chunk = chunker.chunk_file(file)
        audit_log.record({
            "audit_id": audit_id,
            "action": "chunk_created",
            "input": file.path,
            "output": chunk.id,
            "evidence": {"sha256": chunk.sha256}
        })
        
        vector = embedding_service.embed_text(chunk.text)
        audit_log.record({
            "audit_id": audit_id,
            "action": "embedding_created",
            "input": chunk.id,
            "output": {"vector_dim": len(vector)},
            "evidence": {"model": embedding_service.config.model}
        })
        
        store_result = vector_store.add([vector])
        audit_log.record({
            "audit_id": audit_id,
            "action": "vector_stored",
            "input": vector,
            "output": {"id": store_result.id},
            "evidence": {"lancedb_path": str(store_result)}
        })
    
    return {"audit_id": audit_id, "status": "completed"}
```

**Why**: Provides full traceability from document → chunk → vector → search results, matching Jurassic OiO's evidence chain requirement.

---

## Actionable Implementation Plan

### Phase 1: KG1 Instance Resource Graph (Week 1-2)

1. **Define KG1 schema** inspired by Jurassic OiO's IPOMSQ framework
   - Map existing `KnowledgeDocument` metadata to KG1 node types
   - Add `business_semantics` field for domain-specific meaning
   - Add `evidence` field for audit trail

2. **Implement KG1 operations**:
   - `kg1_search(query)` → search in enhanced metadata
   - `kg1_add_node(doc)` → add document to KG1
   - `kg1_relationship(a, b)` → add relationships between nodes

3. **Cross-check**: Ensure KG1 schema integrates with existing `KnowledgeNavigator.topics` and `KnowledgeVectorStore`

### Phase 2: JuraRAG-Enhanced Embedding (Week 2-3)

1. **Modify `EmbeddingService`**:
   - Add `embed_with_context(text, context=None)` method
   - When context provided, prepend to query before Ollama call
   - Add `retrieve_context(text)` method using `KnowledgeNavigator + vector_store`

2. **Add test**: Verify that `embed_with_context()` returns vectors with improved relevance

3. **Cross-check**: Ensure dim=1024 compatibility (per `EmbeddingConfig`)

### Phase 3: KnowledgePlaneService with Maestro-Style Orchestration (Week 3-4)

1. **Create `KnowledgePlaneService`** class:
   - `ingest_directory(path, budget=None)` → end-to-end ingest
   - `search(query, limit, budget_constraints)` → configurable search
   - `health_check()` → Ollama + LanceDB status

2. **Implement configuration optimization** (Maestro-pattern):
   - Multiple embedding configurations (different contexts, chunk sizes)
   - Pareto frontier selection based on budget (tokens, latency, relevance)

3. **Add review workflow** (Jurassic pattern):
   - `ingest_with_review()` → requires approval before storage
   - Evidence logging using existing `AuditLog`

4. **Cross-check**: Run Gate 1 (`pytest -q`, `ruff check`, `mypy src/cyberbrain26`) after each change

### Phase 4: CLI & Documentation (Week 5-6)

1. **Add CLI commands** (per `CAPABILITIES.md`):
   - `cyberbrain knowledge ingest /path/to/docs`
   - `cyberbrain knowledge search "dotaz"`

2. **Update `docs/CAPABILITIES.md`**:
   - Mark knowledge plane capabilities as `WORKING` with evidence `path:line`

3. **Update plan file** with implementation status

---

## Cross-Check Against Codebase

| Component | Existing Code | New Integration | Conflict? |
|-----------|--------------|----------------|-----------|
| `EmbeddingService` | `embedding_service.py:30-48` | Add `embed_with_context()` | ✅ No - extends existing |
| `KnowledgeVectorStore` | `vector_store.py:99-111` | Add KG1-aware search | ✅ No - enhances existing |
| `KnowledgeNavigator` | `navigation.py:104-117` | Add KG1-aware `search_by_keyword()` | ✅ No - extends existing |
| `AuditLog` | `security_plane/audit/chain.py` | Integrate evidence logging | ✅ Yes - requires audit schema update |
| `EmbeddingConfig` | `embedding_config.py:10-37` | Keep dim=1024 standard | ✅ Yes - standard, no change needed |

**Key conflict resolution**: 
- Audit schema update needed but follows existing pattern (`AuditLog.record()` already exists)
- KG1 schema must not break existing `KnowledgeDocument` deserialization → use optional fields

---

## Validation Plan

After each phase, verify:

1. **Phase 1**: `pytest -q` still passes (747 tests), KG1 schema backward-compatible
2. **Phase 2**: `embed_with_context()` returns vectors, relevance test with domain queries
3. **Phase 3**: `KnowledgePlaneService.ingest_directory()` completes end-to-end, Gate 1 green
4. **Phase 4**: CLI commands work, `docs/CAPABILITIES.md` updated with `WORKING` status

**Success criteria**:
- End-to-end pipeline: `.md` file → KG1 node → enhanced embedding → LanceDB store → semantic search → relevant results
- Evidence chain: every operation logged with `audit_id`, traceable from source to result
- Configurable tradeoff: can select between relevance vs. latency vs. cost via Maestro-pattern configs

---

## Open Questions (decision points)

1. **KG1 schema depth**: How many IPOMSQ dimensions to implement? Full 5 (I-P-O-M-S-Q) or subset? **Recommendation**: Start with core 4 (I-P-O-M), add S-Q later.

2. **Context source for RAG**: Pull from KG1 only, or also from `KnowledgeNavigator` topics+tags? **Recommendation**: Both - topics/tags for quick filter, KG1 for deep semantics.

3. **Maestro-pattern complexity**: Full Action Model training + Pareto frontier, or simplified config selection? **Recommendation**: Start with simplified 3-configuration selection (high/medium/low context), add full frontier later.

4. **Audit schema update**: Modify existing `AuditLog` or create separate `KG1AuditLog`? **Recommendation**: Extend existing `AuditLog` with `kg1_operation` field, preserves single source of truth.

---

## Source References

1. **Jurassic Software OiO Launch** (Sept 21, 2026): https://en.jurassic.com.cn/blog/oio-launch/
2. **Huawei CONNECT 2026**: https://www.huawei.com/en/events/huaweiconnect
3. **OiO Product Brief**: https://en.jurassic.com.cn/wp-content/uploads/2026/10/OiO_Product_Brief_EN.pdf
4. **Jurassic OiO Solutions** (May 8, 2026): https://en.jurassic.com.cn/blog/oio-solutions/
5. **White Paper on Next-Gen Intelligent OS** (June 9, 2026): https://en.jurassic.com.cn/blog/next-generation-intelligent-operating-system/
6. **AI21 Labs Maestro Deep Research** (April 28, 2026): https://www.ai21.com/blog/maestro-deep-research-agents/
7. **Jamba2 Release** (Jan 8, 2026): https://www.ai21.com/blog/introducing-jamba2/
8. **AI21 Labs Strategic Pivot** (May 18, 2026): https://olam.business/ai21-labs-the-pivot-to-maestro-after-the-nebius-deal-collapsed
9. **AWS Jurassic Models Pricing** (Oct 8, 2026): https://llmopex.com/changes/1415-aws-now-sells-jurassic-ultra-per-token