TVC Distributed Learning

TVC (Ternary Vector Corpus) is Trinity's distributed continual learning system. It provides zero-forgetting knowledge storage using ternary vector embeddings, bundled into a persistent corpus for instant retrieval.

Commands

tvc-demo

Interactive demonstration of the TVC distributed learning system.

Aliases: tvc

tri tvc-demo

Shows:

• Corpus architecture and configuration
• Hit/miss semantics with cosine similarity
• Bundling and retrieval demo
• Distributed learning flow

Example output:

TVC DISTRIBUTED LEARNING — DEMO
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Architecture:
  ┌──────────────┐    ┌──────────────┐
  │   Query      │───→│  TVC Corpus  │
  │  (embed)     │    │  10K entries │
  └──────────────┘    └──────┬───────┘
                             │
                    ┌────────┴────────┐
                    │  Similarity ≥ φ⁻¹│
                    │  (0.618)?       │
                    └────────┬────────┘
                     YES     │     NO
                 ┌───────┐   │   ┌───────┐
                 │ CACHE  │       │  LLM  │
                 │  HIT   │       │ CALL  │
                 └────────┘       └───────┘

Configuration:
  Capacity:      10,000 entries (100x TextCorpus)
  Dimensions:    1,000 trits per vector
  Threshold:     phi^-1 = 0.618
  Encoding:      Ternary {-1, 0, +1}
  File format:   .tvc (TVC1 magic header)
  Memory:        ~26 MB heap-allocated

tvc-stats

Display TVC corpus configuration and statistics.

tri tvc-stats

Example output:

TVC STATISTICS
═══════════════════════════════════════════════════════════════════
TVC Enabled:       Ready
Max Entries:       10,000
Vector Dimension:  1,000 trits
Threshold:         0.618 (phi^-1)
File Format:       .tvc (TVC1 magic)
═══════════════════════════════════════════════════════════════════

Architecture

How TVC Works

1. Query embedding: User input is converted to a 1,000-trit ternary vector
2. Similarity search: Cosine similarity is computed against all corpus entries
3. Threshold check: If similarity $\geq \phi^{-1} = 0.618$, it's a cache hit
4. Hit: Return cached response instantly (no LLM call needed)
5. Miss: Call LLM, then store the query-response pair in the corpus

Zero-Forgetting Bundling

TVC uses the VSA bundle operation (majority vote) to accumulate all learned patterns into a single memory_vector. This means:

• No patterns are ever discarded
• New knowledge is bundled with existing knowledge
• Retrieval uses cosine similarity against the bundled vector
• Capacity: 10,000 discrete entries + 1 bundled memory vector

Comparison with TextCorpus

Feature	TextCorpus	TVC Corpus
Capacity	100 entries	10,000 entries
Vector type	Float32	Ternary 1
Memory	~40 MB	~26 MB
Compression	None	20x (vs float32)
Forgetting	FIFO eviction	Zero-forgetting
Bundling	No	Yes (majority vote)
File format	JSON	.tvc binary

File Format

TVC files use a binary format with a TVC1 magic header:

Offset  Size    Field
0x00    4       Magic: "TVC1"
0x04    4       Version
0x08    4       Entry count
0x0C    4       Vector dimension
0x10    48      Reserved
0x40    ...     Entries (vector + metadata)

Save path: data/trinity_chat.tvc (auto-saved on REPL exit)

Sacred Mathematics

The TVC threshold is derived from the golden ratio:

$\text{threshold} = \phi^{-1} = \frac{1}{\phi} = \phi - 1 = 0.6180339887...$

This is the Needle threshold (Koschei's Needle) — the minimum similarity required for a cache hit. The number $\phi^{-1}$ is the reciprocal of the golden ratio and appears throughout Trinity's architecture as the quality gate.

Integration with Chat

TVC is automatically loaded when starting the REPL or using tri chat:

# TVC is loaded automatically
tri chat "What is VSA?"

# First time: LLM call + stored in TVC
# Second time: instant cache hit
tri chat "What is VSA?"

The TVC corpus persists across sessions via the data/trinity_chat.tvc file.

See Also

• Interactive REPL — REPL mode with TVC integration
• Core Commands — Chat command details
• Demos & Benchmarks — Needle Check benchmarks
• Sacred Math — Golden ratio constants