Semantic Infrastructure Lab

Version: 1.0
Date: 2025-12-15
Status: Proposal for evaluation
Origin: Session heating-snow-1214

The Core Thesis

The current 7-layer Cognitive OSI Stack was designed product-centric
(where do our 12 projects fit?) rather than problem-centric
(what abstractions solve LLM coexistence?).

If we start from problems instead of products, Provenance emerges as L0, not Philbrick hardware.

The Problem Space

The Semantic OS must solve:

1. Decomposing intent into trackable work - How do we know what we're trying to do?
2. Trust relationships for LLM/AGI coexistence - How do humans and AI work together safely?
3. Cross-domain tooling - How do unrelated domains compose?
4. Meaning manifolds for analogies - How do we find similar things across contexts?

The key insight: Without provenance, you can't trust LLM output.

The Proposed Model

┌─────────────────────────────────────────────────────────────┐
│ L6: REFLECTION                                              │
│ Learning from execution, observability, feedback loops      │
│ Tools: Reveal (structure), metrics, dashboards              │
└─────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────┐
│ L5: EXECUTION                                               │
│ Doing work under constraints, agent orchestration           │
│ Tools: Agent Ether, TIA task execution, schedulers          │
└─────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────┐
│ L4: COMPOSITION                                             │
│ Cross-domain integration, graph routing, topology           │
│ Tools: Pantheon IR, SUP components, workflow engines        │
└─────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────┐
│ L3: INTENT                                                  │
│ What we're accomplishing, contracts, goals, constraints     │
│ Tools: IntentContract, task definitions, acceptance criteria│
└─────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────┐
│ L2: TRUST                                                   │
│ Who can do what, authorization, delegation, capability      │
│ Tools: TAP, AuthorizationGrant, DelegationGrant             │
└─────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────┐
│ L1: MEANING                                                 │
│ Embeddings, types, similarity, semantic units               │
│ Tools: Beth (discovery), Pantheon types, domain schemas     │
└─────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────┐
│ L0: PROVENANCE                                              │
│ Everything has lineage - who made what, when, from what     │
│ Tools: GenesisGraph, hash chains, Merkle trees              │
└─────────────────────────────────────────────────────────────┘

Unix Philosophy Parallel

Unix insight: "Everything is a file"
- Files are the universal abstraction
- Tools compose through file descriptors
- The filesystem is the namespace

Semantic OS insight: "Everything has meaning and provenance"
- Semantic objects are the universal abstraction
- Tools compose through typed interfaces
- The provenance graph is the trust foundation

Fundamental Primitives

Unix lacked provenance - you could copy a file, modify it, and lose all history. The Semantic OS makes lineage a first-class concern.

How Tools Fit (Spanning Layers)

Tools are like Unix utilities - they span layers based on function, not architecture:

This is why component assignment has been chaotic - tools naturally span layers.

Why Provenance at L0?

The Trust Problem

When an LLM generates code, documentation, or analysis:
- Who prompted it?
- What context did it have?
- What model version?
- What transformations occurred?

Without provenance, you can't answer these questions. Without answers, you can't trust the output.

The Coexistence Problem

Human-AI collaboration requires:
1. Knowing what came from where
2. Verifying transformations
3. Attributing decisions
4. Rolling back changes

All of these depend on provenance.

The Philbrick Problem

The current model puts Philbrick hardware at L0 because:
- It's physical
- It's foundational to computing

But Philbrick is:
- A fun hardware project
- Not essential for modern LLM architectures
- Not solving the trust/coexistence problem

Philbrick becomes an optional backend, not the foundation.

Layer Dependencies

L6 Reflection   ← needs execution results from L5
     ↑
L5 Execution    ← operates within composition from L4
     ↑
L4 Composition  ← composes elements with intent from L3
     ↑
L3 Intent       ← intent authorized by trust from L2
     ↑
L2 Trust        ← trust based on meaning from L1
     ↑
L1 Meaning      ← meaning grounded in provenance from L0
     ↑
L0 Provenance   ← foundation: everything has lineage

Each layer depends on the layer below. This dependency chain makes sense:
- You can't trust without knowing meaning
- You can't have intent without trust
- You can't compose without intent
- You can't execute without composition
- You can't reflect without execution

Comparison with Current Model

Implications

For GenesisGraph

Becomes foundational infrastructure, not just "provenance tooling." Every operation in the Semantic OS should create provenance records.

For Trust/Authorization

TAP and the Hierarchical Agency Framework become L2 infrastructure, not just "intent-layer features."

For Agent Development

Agents operate at L5 (Execution) but must:
- Respect L2 (Trust) constraints
- Express L3 (Intent) contracts
- Produce L6 (Reflection) observability

For New Projects

When adding a project, ask:
1. What layer does it primarily serve?
2. What layers does it span?
3. How does it relate to provenance?

Open Questions (Addressed)

1. Is this too abstract? Does problem-centric lose implementation guidance?

Answer: No - it provides better guidance.

The brewing-sleet synthesis found "we have more implemented than documented, but critical Layer 1 primitives remain unbuilt." The problem-centric framing tells us what to build (Intent Verification, Uncertainty Tracking, Cross-Domain Translation) rather than just where to put things.

Implementation guidance comes from the layer dependencies:
- Build L0 (GenesisGraph provenance) first
- Then L1 (Beth meaning/types)
- Then L2 (TAP trust infrastructure)
- Then L3-L6 as needed

2. Where does memory fit? The Original model had Semantic Memory at L0.

Answer: Memory is implementation, not abstraction.

Semantic memory (storage, indexing) is how we implement L1 (Meaning) and L0 (Provenance). It's infrastructure, not architecture. Just as TCP/IP doesn't have a "RAM layer," the Semantic OS doesn't need memory as a layer - it's assumed.

Beth's index, Gemma's storage, and S3/database backends are all L1 implementation details.

3. Are 7 layers right? Could Trust and Meaning collapse?

Answer: 7 layers is defensible; collapsing would lose important distinctions.

• Trust ≠ Meaning: You can understand what something means without being authorized to act on it
• Execution ≠ Reflection: Doing work and learning from work are different concerns

The OSI model has 7 layers. The Cognitive OSI has 7+meta. 7 seems to be a natural granularity for layered architectures.

4. What about hardware? If Philbrick isn't L0, where does physical computing fit?

Answer: L-1 (Substrate) - below the semantic layers.

The recommended hybrid adds:

L0: Provenance       - Everything has lineage
─────────────────────────────────────────────
L-1: Substrate       - Physical/computational reality (Philbrick, optional)

This preserves Philbrick's place without making hardware the semantic foundation. The Semantic OS runs on hardware but isn't defined by it.

5. How do we migrate? Can we align existing docs incrementally?

Answer: Yes, with a phased approach.

1. Phase 1: Accept this model as canonical (stakeholder review)
2. Phase 2: Update SIL_GLOSSARY.md with new layer definitions
3. Phase 3: Add frontmatter to existing docs noting which model they use
4. Phase 4: Incrementally update canonical docs (SEMANTIC_OS_ARCHITECTURE, SEMANTIC_OBSERVABILITY, etc.)
5. Phase 5: Archive historical models with "superseded by" notes

Evaluation Criteria

Score this model against:

Weighted Total: 4.15 (highest of all 5 models evaluated)

See MODEL_EVALUATION.md for full comparison and rationale.

References

• sessions/heating-snow-1214/README_2025-12-15_08-52.md - Origin of proposal
• K&R/Unix Philosophy documentation
• GenesisGraph project documentation
• Trust Assertion Protocol specification

Synthesis: Sessions That Informed This Model

This proposal synthesizes insights from 6+ sessions:

Key Synthesis Insights

1. "We have more implemented than documented" (brewing-sleet)
   - Infrastructure exists; Layer 1 primitives missing
   - Problem isn't building from scratch; it's formalizing and connecting

2. "Layer 1 is the glue" (brewing-sleet)
   - Intent Verification, Uncertainty Tracking, Cross-Domain Translation
   - These connect Agent coordination (L3) ↔ Pantheon IR (L1) ↔ Memory (L0)

3. "4 competing models" (enchanted-centaur)
   - Models evolved independently in different docs
   - No single authoritative version until now

4. "Provenance when I hear SOS" (Scott, heating-snow)
   - Founder intuition aligned with problem-centric analysis
   - Trust and lineage are core to the SIL mission

The Case for Adoption

This model should become canonical because:

1. Highest evaluation score (4.15 vs 3.05 for Canonical)
2. Resolves component assignment chaos (tools span layers)
3. Addresses core SIL mission (LLM coexistence requires trust)
4. Founder alignment (Scott's instinct matches the analysis)
5. Explains historical divergence (product-centric vs problem-centric)

Status

Proposed for adoption. Awaiting stakeholder review before updating canonical docs.