SNAG vs RAG

The Fundamental Problem

Large Language Models are powerful storytellers, but they’re fragile reasoners. Naive single-prompt simulations collapse beyond ~10 entities or ~20 interactions due to:

Inconsistency drift: Entities forget facts, relationships change arbitrarily
Token explosion: Full-context approaches hit limits fast
Causal incoherence: No tracking of who knows what, when they learned it
Hallucinated knowledge: Entities magically know things they couldn’t

What is SNAG?

SNAG (Social Network Augmented Generation) is to social simulation what RAG (Retrieval Augmented Generation) is to document search.

RAG

Retrieves documents to ground generation in factual knowledge

SNAG

Synthesizes social graphs to ground generation in causal structure

SNAG vs RAG: Side-by-Side

Dimension	RAG	SNAG (Timepoint Pro)
Grounds LLMs in	Retrieved documents	Synthesized social graphs
Maintains	Document relevance	Causal provenance + temporal consistency
Scales to	Millions of documents	Dozens of entities, hundreds of timepoints
Output	Grounded answers	Auditable causal simulations + training data
Core Structure	Document embeddings + retrieval	Entity tensors + exposure events + causal chains
Validation	Relevance scoring	Multi-constraint validation (information conservation, energy budgets, network flow)

How SNAG Grounds LLMs

Every entity exists in a typed graph with:

Entity:
  - entity_id: str
  - entity_type: "human" | "animal" | "building" | "abstract"
  - tensor: TTMTensor  # Context, biology, behavior vectors
  - resolution_level: ResolutionLevel  # TENSOR_ONLY → TRAINED
  - entity_metadata: Dict  # Cognitive + physical state

2. Causal Provenance

Knowledge doesn’t appear magically—it has exposure events:

ExposureEvent:
  - entity_id: "thomas_jefferson"
  - event_type: "learned" | "witnessed" | "told"
  - information: "Madison's Virginia Plan notes"
  - source: "james_madison"
  - timestamp: datetime
  - confidence: 0.95
  - timepoint_id: "constitutional_convention_day_3"

Validation constraint: An entity cannot know something without a recorded exposure event explaining how they learned it.

Example: Constitutional Convention Knowledge Flow

Day 1: Madison creates Virginia Plan → exposure event (type=created, source=self)Day 3: Madison shares with Washington → exposure event (type=told, source=madison, entity=washington)Day 5: Washington references plan in debate ✅ Valid (has exposure from Day 3)Day 5: Jefferson references plan ❌ INVALID (no exposure event, Jefferson not present)

3. Temporal Consistency

Timepoints form explicit causal chains:

Timepoint:
  - timepoint_id: "board_meeting_q3_2024"
  - causal_parent: "board_meeting_q2_2024"  # Explicit link
  - event_description: "CFO presents revised forecast"
  - entities_present: ["ceo", "cfo", "vp_eng"]
  - timestamp: datetime

An entity at timepoint T can only reference information from timepoints T’ where a causal path exists from T’ → T.

4. Variable-Depth Fidelity

Not all entities deserve equal computational attention:

TENSOR_ONLY (~200 tokens)

Background entities: compressed state vectors only

SCENE (~2k tokens)

Minor participants: scene-level behavior

GRAPH (~5k tokens)

Secondary characters: relationship tracking

DIALOG (~10k tokens)

Key participants: full dialog generation

TRAINED (~50k tokens)

Protagonists: complete psychological depth

Fidelity is query-driven: entities start compressed and elevate on-demand.

Why This Matters

Transform LLMs from Storytellers to Reasoners

SNAG’s structured propagation, variable-depth fidelity, and composable mechanisms let you scale to dozens of entities across hundreds of timepoints while keeping costs low and causality auditable.

Exponential Value at Scale

The larger and more intricate the social system, the more emergent behaviors surface that intuition misses:

# Board + investors + competitors
# 12 entities, 24 timepoints, 3 counterfactual branches
# Cost: ~$0.35, Time: ~8 minutes
./run.sh run vc_pitch_branching

Superior Training Data

SNAG outputs include:

Full causal ancestry: Every fact has provenance
Counterfactual branches: “What if” scenarios with controlled interventions
Quantitative states: Emotion tensors, energy budgets, confidence levels
Dialog with context: Every conversation includes relationship state, knowledge access, emotional tone

The 95% Cost Reduction

Traditional approach: uniform high fidelity for all entities

100 entities × 10 timepoints × 50k tokens = 50M tokens

SNAG approach: heterogeneous fidelity (power-law distribution)

~10% TRAINED (5M tokens)
~20% DIALOG (2M tokens)
~30% SCENE-GRAPH (1.5M tokens)
~40% TENSOR_ONLY (80k tokens)
Total: ~2.5M tokens (95% reduction)

Actual costs: $0.15–$ 1.00 per run depending on complexity and temporal mode.

Architecture Insight

SNAG treats fidelity as a query-driven 2D surface over (entity, timepoint) space:

          Timepoints →
Entities  T0   T1   T2   T3   T4
  ↓
CEO       [D] [T] [T] [T] [D]   ← High centrality
CFO       [D] [D] [G] [S] [S]   ← Financial pivots
Attendee  [S] [S] [S] [S] [S]   ← Background only

T=TRAINED, D=DIALOG, G=GRAPH, S=SCENE

Resolution is mutable: queries elevate detail on-demand while preserving all established causal constraints.

Key Mechanisms

M1: Heterogeneous Fidelity

Power-law resolution distribution

M3: Exposure Events

Tracked knowledge acquisition

M7: Causal Chains

Explicit temporal ancestry

M6: TTM Compression

97% compression with structure

M11: Dialog Synthesis

Per-character turn generation

M17: Temporal Modes

5 causality regimes

Next Steps

Temporal Modes

Learn how time itself can have different causal rules

Fidelity Management

Deep dive into resolution levels and TTM tensors

Knowledge Provenance

How exposure events prevent anachronisms

All 19 Mechanisms

Complete technical architecture

Getting Started

Core Concepts

Temporal Modes

Mechanisms

Guides

Examples

The Fundamental Problem

What is SNAG?

RAG

SNAG

SNAG vs RAG: Side-by-Side

How SNAG Grounds LLMs

2. Causal Provenance

3. Temporal Consistency

4. Variable-Depth Fidelity

Why This Matters

Transform LLMs from Storytellers to Reasoners

Exponential Value at Scale

Superior Training Data

The 95% Cost Reduction

Architecture Insight

Key Mechanisms

M1: Heterogeneous Fidelity

M3: Exposure Events

M7: Causal Chains

M6: TTM Compression

M11: Dialog Synthesis

M17: Temporal Modes

Next Steps

Temporal Modes

Fidelity Management

Knowledge Provenance

All 19 Mechanisms

Build docs developers (and LLMs) love

Getting Started

Core Concepts

Temporal Modes

Mechanisms

Guides

Examples

​The Fundamental Problem

​What is SNAG?

RAG

SNAG

​SNAG vs RAG: Side-by-Side

​How SNAG Grounds LLMs

​1. Structured Social Graph

​2. Causal Provenance

​3. Temporal Consistency

​4. Variable-Depth Fidelity

​Why This Matters

​Transform LLMs from Storytellers to Reasoners

​Exponential Value at Scale

​Superior Training Data

​The 95% Cost Reduction

​Architecture Insight

​Key Mechanisms

M1: Heterogeneous Fidelity

M3: Exposure Events

M7: Causal Chains

M6: TTM Compression

M11: Dialog Synthesis

M17: Temporal Modes

​Next Steps

Temporal Modes

Fidelity Management

Knowledge Provenance

All 19 Mechanisms

Build docs developers (and LLMs) love

The Fundamental Problem

What is SNAG?

SNAG vs RAG: Side-by-Side

How SNAG Grounds LLMs

1. Structured Social Graph

2. Causal Provenance

3. Temporal Consistency

4. Variable-Depth Fidelity

Why This Matters

Transform LLMs from Storytellers to Reasoners

Exponential Value at Scale

Superior Training Data

The 95% Cost Reduction

Architecture Insight

Key Mechanisms

Next Steps