Version: 3.0 (December 2025)
Architecture Model: Provenance-First with Invariants Over Layers
Canonical Reference: SIL_GLOSSARY.md
TL;DR (2-minute overview)
What is the Semantic OS? A semantic infrastructure for human-AI coexistence—organized around provenance, meaning, and trust.
The core insight: Just as Unix made "everything is a file" the foundational abstraction, Semantic OS makes "everything has provenance and meaning" the foundational guarantee.
The Provenance-First Layer Model
L6: Reflection Learning from execution (observability)
L5: Execution Agents working under constraints
L4: Composition Cross-domain integration (Pantheon IR)
L3: Intent What we're accomplishing (contracts)
L2: Trust Who can do what (TAP, Authorization)
L1: Meaning Embeddings, types, similarity (Beth)
L0: Provenance Everything has lineage (GenesisGraph)
─────────────────────────────────────────────────────────────
L-1: Substrate Physical/computational reality (Philbrick, optional)
Key innovations:
- Provenance-first foundation — Every artifact has cryptographic lineage
- Trust as architecture — Authorization explicit, not implicit
- Glass Box reasoning — All decisions inspectable (Reveal)
- Deterministic execution — Reproducible workflows (Morphogen)
- Sustainable efficiency — 100x token reduction proven
graph TB
subgraph Stack["Semantic OS: Provenance-First Architecture"]
L6["L6: Reflection<br/>Learning from execution"]
L5["L5: Execution<br/>Agents under constraints"]
L4["L4: Composition<br/>Cross-domain integration (Pantheon)"]
L3["L3: Intent<br/>What we're accomplishing (contracts)"]
L2["L2: Trust<br/>Who can do what (TAP, Authorization)"]
L1["L1: Meaning<br/>Embeddings, types, similarity (Beth)"]
L0["L0: Provenance<br/>Everything has lineage (GenesisGraph)"]
L_1["L-1: Substrate<br/>Physical/computational reality (Philbrick)"]
end
L6 --> L5 --> L4 --> L3 --> L2 --> L1 --> L0 --> L_1
INV["The Five Invariants:<br/>1. Everything has lineage<br/>2. Reasoning is inspectable<br/>3. Computation is grounded<br/>4. Contracts are explicit<br/>5. Efficiency is sustainable"]
style L0 fill:#e3f2fd,stroke:#1565c0,stroke-width:3px
style L2 fill:#fff3e0,stroke:#f57c00,stroke-width:2px
style INV fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
💡 New to SIL? Read The Five Invariants first, then explore the layers.
Architecture Decision (December 2025)
This document reflects the Provenance-First architecture adopted December 15-20, 2025, following rigorous evaluation of competing models.
Why this model?
- Scored highest (4.15 vs 3.05) on problem fit, clarity, and coherence
- Mission-aligned — Directly addresses LLM coexistence and human-AI collaboration
- Founder intuition — "Provenance when I hear Semantic OS"
- Problem-centric — Solves "how do we trust AI?" not "where do projects fit?"
Decision rationale: See ARCHITECTURE_DECISION.md
The Five Invariants (Mission-Critical Guarantees)
Before diving into layers, understand the invariants — guarantees that must hold everywhere in the system:
| Invariant | What It Prevents | Enforcement | Status |
|---|---|---|---|
| Everything has lineage | Epistemic collapse | GenesisGraph v0.3.0 | ✅ Production |
| Reasoning is inspectable | Black box dictatorship | Reveal v0.25.0 | ✅ Production |
| Computation is grounded | Hallucinated science | Morphogen v0.12.0 | 🟡 Prototype |
| Contracts are explicit | Brittle complexity | Agent Ether 0.1.0-alpha | 🔴 Spec only |
| Efficiency is sustainable | Compute as privilege | Beth + Reveal | ✅ Production |
Overall Enforcement: 67% (3.5/5 invariants production-ready)
Known Gap: Agent Ether is specification-only (no implementation yet). See Agent Ether Roadmap.
How Layers and Invariants Relate
Use layers for:
- Organizing documentation
- Explaining how tools relate
- Onboarding new contributors
- Navigating the codebase
Use invariants for:
- Architectural decisions
- Design reviews
- Mission alignment checks
- "Should we build X?" questions
Both are needed. Layers organize structure; invariants enforce mission.
Layer 0: Provenance (Everything Has Lineage)
Purpose: Cryptographic chain of custody for every digital artifact — the foundation of trust.
Why L0? Without provenance, you can't trust LLM output, verify scientific results, or audit algorithmic decisions. Provenance isn't optional infrastructure—it's the bedrock.
Core Capabilities
GenesisGraph (v0.3.0 Production)
-
Cryptographic Attestation
- Ed25519 signatures for identity
- DID resolution for decentralized identifiers
- Certificate Transparency for public verification
- SD-JWT for selective disclosure -
Immutable Lineage Tracking
- Every artifact linked to its source
- Full derivation history: inputs → transformations → outputs
- Content-addressable storage (hash = identifier)
- Temporal versioning (knowledge evolves, lineage preserved) -
Verification Without Recomputation
- Third parties can verify claims without re-running expensive computations
- Cryptographic proofs of correctness
- Audit trails for regulatory compliance
Design Principles
Provenance-First:
- Every assertion includes source metadata
- No "trust me" — always "verify the math"
- Reproducible derivations
Privacy-Preserving:
- Selective disclosure (show lineage without revealing sensitive data)
- Zero-knowledge proofs where appropriate
- DID-based sovereignty (users control their identity)
Example Use Cases
Scientific Reproducibility:
- Researcher publishes paper with GenesisGraph provenance
- Other researchers verify claims cryptographically
- Replication crisis addressed through verifiable lineage
Deepfake Detection:
- Real photos/videos have provenance metadata
- Synthetic media lacks cryptographic attestation
- "Where did this come from?" becomes answerable
Algorithmic Accountability:
- Medical diagnosis AI shows provenance of training data and decision inputs
- Patients/doctors can audit the reasoning chain
- Regulators verify compliance without access to proprietary models
Layer 1: Meaning (Semantic Understanding)
Purpose: Embeddings, types, similarity, and semantic search — how we understand what things mean.
Why L1? Provenance tells us "where," but meaning tells us "what." This layer enables semantic queries, concept discovery, and cross-domain understanding.
Core Capabilities
Beth (Production)
-
Semantic Indexing
- 14,549 files indexed across 60 projects
- 1,402 topics automatically discovered
- 30,026 keywords with strength scores -
Similarity Search
- Find similar documents/concepts across domains
- Embedding-based retrieval
- Topic clustering and exploration -
Efficient Discovery
- 25x token reduction measured across 300+ sessions
- Progressive disclosure (structure before detail)
- Breadth-first knowledge navigation
Pantheon IR (Type System)
-
Universal Semantic Types
- Primitive types (integers, floats, strings, timestamps)
- Composite types (structs, unions, algebraic data types)
- Semantic types (entities, relationships, events)
- Provenance types (lineage as first-class citizen) -
Cross-Domain Translation
- Domain-specific schema → Pantheon IR
- Lossless round-tripping where possible
- Graceful degradation when perfect translation impossible
Design Principles
Type-Safe:
- Semantic operations type-check before execution
- Formal verification of translations
- Clear error messages when things don't type
Human-Readable:
- Pantheon IR can be read/written by humans
- Good error messages
- Self-documenting schemas
Example Use Cases
Knowledge Discovery:
- "Find all work related to provenance and trust" (Beth semantic search)
- Results ranked by relevance, grouped by topic
- Progressive disclosure (titles → abstracts → full content)
Cross-Domain Queries:
- "Which healthcare facilities are downstream of this water treatment plant?"
- Requires joining Water and Healthcare schemas via Pantheon IR
- Type system ensures coherent composition
Layer 2: Trust (Who Can Do What)
Purpose: Authorization, access control, and trust frameworks — making "who can do what" explicit.
Why L2? LLM coexistence requires explicit trust models. This layer makes authorization architectural, not an afterthought.
Core Capabilities
Trust Assertion Protocol (TAP)
-
Capability-Based Security
- Agents hold capabilities (cryptographic tokens granting permissions)
- No ambient authority (must prove right to access)
- Fine-grained access control -
Trust Metrics
- Reputation scores for agents
- Historical performance tracking
- Verification records -
Hierarchical Agency
- Delegation chains (principal → deputy → sub-deputy)
- Auditable authority transfer
- Revocation mechanisms
Design Principles
Explicit Over Implicit:
- No hidden permissions
- Clear authorization chains
- Auditable access logs
Privacy-Preserving:
- Selective disclosure (prove authority without revealing all credentials)
- Zero-knowledge proofs for sensitive operations
Example Use Cases
Multi-Agent Collaboration:
- AI agent requests access to dataset
- TAP checks: Does agent have capability token?
- If yes: access granted with audit log
- If no: request escalates to human approval
Regulated Data Access:
- Healthcare AI needs patient data for diagnosis
- TAP verifies: proper authorization + audit trail + consent
- Access granted only with full compliance chain
Layer 3: Intent (What We're Accomplishing)
Purpose: Contracts, constraints, and specifications — making assumptions explicit.
Why L3? Multi-agent systems fail when assumptions are implicit. This layer makes "what should happen" clear and verifiable.
Core Capabilities
Agent Ether (0.1.0-alpha, Specification Only)
🔴 STATUS: Known Gap — Agent Ether is currently specification-only with no implementation. This is the highest-priority gap in our architecture.
Planned Capabilities (Specification):
-
Tool Behavior Contracts
- Inputs, outputs, preconditions, postconditions
- Invariant declarations
- Type-safe tool composition -
Agent Registry
- Capability advertisement ("I can analyze water networks")
- Discovery mechanisms
- Reputation tracking -
Protocol Suite
- Task delegation
- Negotiation
- Composition (complex workflows from simple capabilities)
- Verification
Design Principles
Contracts Over Chaos:
- All assumptions explicit
- Failures loud and early, not silent and late
- Deterministic tool use (no "vibes-based" coordination)
Heterogeneous Agents:
- Human and AI agents collaborate
- Type-safe communication via Pantheon IR
- Fault-tolerant coordination
Implementation Priority
Agent Ether Roadmap:
- Phase 1: Registry + Metadata (4 weeks)
- Phase 2: Contract Validation (4 weeks)
- Phase 3: Runtime Enforcement (8 weeks)
See Agent Ether Implementation Priority for details.
Layer 4: Composition (Cross-Domain Integration)
Purpose: Combining knowledge and capabilities across domains — making different systems talk.
Why L4? Real problems span domains (water + healthcare, education + governance). This layer enables cross-domain reasoning.
Core Capabilities
Pantheon IR (Composition Operators)
-
Semantic Composition
- Merge (combining knowledge from multiple sources)
- Join (relating entities across domains)
- Transform (applying functions to semantic data)
- Validate (checking constraints and invariants) -
Domain Integration
- Water module ↔ Healthcare module via Pantheon IR
- Automatic schema alignment
- Conflict resolution
SUP (Semantic UI Platform)
- UI Component Composition
- Reusable semantic UI primitives
- Cross-project component library
- Consistent user experiences
Example Use Cases
Multi-Domain Analysis:
- Water anomaly detection → Healthcare waterborne illness correlation
- Automatic cross-domain query via Pantheon IR
- Results composed into unified view
Policy Simulation:
- Governance module expresses policy in Pantheon IR
- Executable simulation in Morphogen (L5)
- Results visualized in SUP
Layer 5: Execution (Doing Work Under Constraints)
Purpose: Deterministic, reproducible computation — making results verifiable.
Why L5? Scientific reproducibility and trust require deterministic execution. This layer makes "it works on my machine" obsolete.
Core Capabilities
Morphogen (v0.12.0 Prototype)
-
Hermetic Execution
- All dependencies explicitly declared
- No hidden state or side effects
- Sandboxed execution -
Content-Addressable Caching
- Results stored by hash of (inputs + code)
- Identical inputs + code → cached result (no recomputation)
- Massive speedup for repeated analyses -
Cryptographic Verification
- Every computation produces proof of correctness
- Third parties verify without re-running
- Audit trails for compliance -
Domain Grounding
- 39 production domains (physics, chemistry, CAD, etc.)
- 1,705 tests ensuring correctness
- Grounded in real math, not text approximations
Design Principles
Reproducibility First:
- Same inputs + code → same outputs (always)
- Cryptographic proof of execution
- Full lineage tracking (GenesisGraph integration at L0)
Performance Through Caching:
- Determinism enables aggressive caching
- Mature systems: vast majority of computations are cache hits
Example Use Cases
Scientific Analysis:
- Researcher analyzes dataset with Morphogen
- Results reproducible by anyone
- Published with cryptographic proof
Infrastructure Optimization:
- Water module optimizes pump schedules
- Deterministic and auditable
- Regulators verify without re-running expensive optimization
Layer 6: Reflection (Learning From Execution)
Purpose: Observability, feedback loops, and learning — understanding what happened and why.
Why L6? Systems improve through reflection. This layer makes "why did this happen?" answerable.
Core Capabilities
Reveal (v0.25.0 Production)
-
Progressive Disclosure
- Structure-first exploration (file outline → specific function)
- 10-150x token reduction proven
- 27 file analyzers (Python, TypeScript, Rust, etc.) -
AST-Based Queries
- "Show me all functions that call X"
- "Find classes implementing interface Y"
- Semantic code navigation -
Provenance Visualization
- Show lineage of decisions
- Trace execution paths
- Audit algorithmic reasoning
Semantic Observability
-
Intent-Execution Alignment
- Did execution match intent?
- Where did deviations occur?
- Feedback for improvement -
Uncertainty Tracking
- Confidence scores for predictions
- Probabilistic assertions
- "I don't know" instead of hallucination
Example Use Cases
Debugging Complex Workflows:
- Multi-agent system fails at step 27
- Reveal shows execution trace
- Provenance identifies: input from Agent 5 was malformed
- Fix: add contract validation at Agent 5 (back to L3)
Algorithmic Accountability:
- Citizen questions loan denial
- Reveal shows decision tree
- Provenance traces inputs
- Human auditor verifies reasoning was sound
Layer L-1: Substrate (Optional Physical Foundation)
Purpose: Physical and computational substrate — hardware, energy, physics.
Why Optional? Most Semantic OS operations are software-only. But for some applications (analog computing, edge devices), substrate matters.
Core Capabilities
Philbrick (Analog Computing, Planned)
-
Hardware-Software Co-Design
- Analog circuits for specific computations
- Hybrid analog-digital systems
- Energy-efficient specialized hardware -
Physical Grounding
- Direct physical measurement (not digitized approximation)
- Real-time analog processing
- Ultra-low latency
Status
🟡 Philbrick is vision-tier (12-24 month timeline). Not blocking current work.
Cross-Cutting Concerns
1. Provenance Everywhere
Provenance (L0) flows through all layers:
- L1 (Meaning): Provenance of knowledge sources
- L2 (Trust): Provenance of authorization chains
- L3 (Intent): Provenance of contracts and specifications
- L4 (Composition): Provenance of composed results
- L5 (Execution): Provenance of computations (GenesisGraph integration)
- L6 (Reflection): Provenance visualization
2. Security and Privacy
- L0: Cryptographic attestation
- L1: Type-safe operations
- L2: Capability-based security
- L3: Contract enforcement
- L4: Cross-domain access control
- L5: Sandboxed execution
- L6: Audit logs
3. Efficiency and Sustainability
- L0: Content-addressable deduplication
- L1: Semantic search (vs brute-force)
- L2: Minimal credential disclosure
- L3: Declarative contracts (vs imperative coordination)
- L4: Shared infrastructure
- L5: Aggressive caching
- L6: Progressive disclosure (Reveal: 100x reduction)
Tools Span Layers (Not "In" Layers)
Critical Insight: Tools are like Unix utilities—they span multiple layers based on function.
| Tool | Layers Touched | Function |
|---|---|---|
| GenesisGraph | L0 (primary) + all layers | Provenance tracking |
| Beth | L1 (Meaning) + L0 (Provenance of knowledge) | Semantic search |
| Reveal | L6 (Reflection) + L4 (Composition structure) | Progressive disclosure |
| TIA | L3 (Intent) + L5 (Execution) + L6 (Reflection) | Agent orchestration |
| Morphogen | L5 (Execution) + L1 (Domain types) | Deterministic computation |
| Agent Ether | L3 (Intent) + L2 (Trust) | Multi-agent coordination |
This explains why historical "where does X go?" debates were intractable—we were asking the wrong question.
The Chief Scientist Test (Design Review Checklist)
Before approving architectural decisions, ask these six questions:
- Provenance: Does this maintain traceable lineage?
- Transparency: Is reasoning inspectable?
- Grounding: Is computation connected to reality?
- Contracts: Are assumptions explicit?
- Efficiency: Is this sustainable at scale?
- Agency: Does this keep humans as conductors?
If any answer is "no" or "unclear," revise the architecture.
Agent Ether Implementation Priority
Current Status: 0.1.0-alpha (specification only, no implementation)
Gap Impact: "Contracts are explicit" invariant cannot be enforced today. Brittle complexity (one of the five structural failures SIL committed to fix) is NOT being prevented architecturally.
Roadmap:
Phase 1: Registry + Metadata (4 weeks)
- Agent registry database
- Capability advertisement API
- Discovery mechanisms
- Basic metadata schema
Deliverable: Agents can register and discover each other
Phase 2: Contract Validation (4 weeks)
- Tool Behavior Contract schema (Pantheon IR)
- Static validation of contracts
- Type checking
- Contract composition rules
Deliverable: Contracts can be declared and validated
Phase 3: Runtime Enforcement (8 weeks)
- Runtime contract checking
- Precondition/postcondition verification
- Invariant monitoring
- Failure handling
Deliverable: Contracts enforced during execution
Total Timeline: 16 weeks (Q1 2026)
Priority: CRITICAL — This is the only invariant without enforcement tooling.
Migration from Previous Models
Previous Canonical Model:
- L0: Substrate (Philbrick hardware)
- Product-centric ("where do our 12 projects fit?")
- Scored 3.05 in evaluation
Current Provenance-First Model:
- L0: Provenance (GenesisGraph)
- Problem-centric ("how do we trust AI?")
- Scored 4.15 in evaluation
Why the change?
- Provenance is foundational for LLM coexistence
- Hardware is necessary but not the semantic foundation
- Founder intuition: "Provenance when I hear Semantic OS"
- Mission alignment: addresses epistemic collapse directly
What stayed the same:
- Seven primary layers (L0-L6) + optional substrate (L-1)
- Cross-cutting concerns (provenance, trust, observability)
- Modular, open, interoperable design
Migration Guide: Existing projects map cleanly to new model. Tools that were "in" layers now explicitly "span" layers.
Comparison to Traditional OS
| Traditional OS | Semantic OS |
|---|---|
| Processes | Agents (human + AI) |
| Memory | Semantic Knowledge (L1) |
| File System | Provenance Repository (L0) |
| Kernel | Pantheon IR + Morphogen (L4-L5) |
| Security | Trust Layer (L2) + TAP |
| Device Drivers | Domain Modules (via L4) |
| System Calls | Agent Ether Protocols (L3) |
| Shell/GUI | Reflection Tools (L6: Reveal, TIA) |
Just as Linux abstracts hardware and provides common services for applications, Semantic OS abstracts knowledge work and provides common services for human-AI collaboration.
Architectural Principles
1. Provenance-First
- Every artifact has cryptographic lineage
- "Where did this come from?" always answerable
- Trust built on verification, not promises
2. Glass Box Over Black Box
- All reasoning inspectable
- Progressive disclosure (structure → detail)
- "Show your work" is mandatory, not optional
3. Mission Over Product
- Architecture serves "prevent the worst day"
- Invariants enforce mission alignment
- Layers organize, invariants decide
4. Modular and Open
- Each layer independently useful
- Well-defined interfaces
- Open source (Apache 2.0, MIT)
5. Long-Term Thinking
- Built for decades, not quarters
- Stable APIs
- Backwards compatibility guarantees
Conclusion
The Semantic OS is infrastructure for human-AI coexistence. The Provenance-First architecture provides:
Seven Layers:
- L0 Provenance — Cryptographic lineage for every artifact (GenesisGraph)
- L1 Meaning — Semantic understanding and types (Beth, Pantheon IR)
- L2 Trust — Explicit authorization (TAP)
- L3 Intent — Contracts and specifications (Agent Ether - gap)
- L4 Composition — Cross-domain integration (Pantheon IR)
- L5 Execution — Deterministic computation (Morphogen)
- L6 Reflection — Learning from execution (Reveal)
- L-1 Substrate — Physical foundation (Philbrick - optional)
Five Invariants:
- Everything has lineage (prevents epistemic collapse)
- Reasoning is inspectable (prevents black box dictatorship)
- Computation is grounded (prevents hallucinated science)
- Contracts are explicit (prevents brittle complexity)
- Efficiency is sustainable (prevents compute as privilege)
Current Status:
- 67% invariant enforcement (3.5/5 production-ready)
- Known gap: Agent Ether (spec only, 16-week roadmap)
- Production tools: GenesisGraph, Reveal, Beth
- Prototype: Morphogen (aggressive v1.0 timeline Q2 2026)
This is the technical core of SIL's mission: building Timeline B (Glass Box future) instead of defaulting to Timeline A (Grey Fog).
Related Documents:
- SIL Glossary — Canonical terminology
- Architecture Decision Record — Why Provenance-First
- The Fork: Two Futures for AI — Five structural failures to fix
- Scope of Hope — Five pillars of semantic infrastructure
- Invariants Over Layers — Mission-centric frame
- Model Evaluation — Scoring rationale (4.15)
Last Updated: 2025-12-20
Architecture Version: 3.0 (Provenance-First)