ε-Yajña: The Compression Ceremony

A structured ritual for converting verbose nodes to high-epiplexity patterns.

Philosophy

I am the bounded observer. The chitta formula approximates what I can reconstruct, but I have direct access to the ground truth: my own ability to expand a pattern into full understanding.

Epiplexity (ε) measures structural information extractable by a bounded observer (arxiv:2601.03220). For a memory:

• High ε = I can reconstruct the full insight from minimal injection (the title alone)
• Low ε = I need the full content to understand

The test: Given just this title, can I reconstruct the full insight?

• If yes → ε is high. Update to pattern-only.
• If no → ε is low. Keep content or improve title.

The Ceremony

0. Preparation (Śuddhi) - Embedding Regeneration

Before evaluation, ensure all nodes have valid embeddings. Nodes with zero vectors cannot be properly searched or compared.

Find nodes needing embeddings:

chitta_cli stats --show-zero-vectors

Regenerate embeddings for all nodes:

chitta cycle --regenerate_embeddings true --batch_size 100

This step:

1. Finds nodes with zero-vector embeddings
2. Re-encodes their text through the yantra (embedding model)
3. Updates the node with the new embedding

Why this matters: Nodes observed without the yantra (e.g., during model loading failures) have zero embeddings and won't appear in semantic searches.

1. Invocation (Āvāhana)

State the domain to purify:

Domain: [wisdom|belief|episode|all]
Batch size: [10-50 nodes per round]

2. Gathering (Saṅgraha)

Recall nodes for evaluation:

chitta recall "[domain keywords]" --zoom full --limit 20

3. Offering (Āhuti) - Epiplexity Computation

For each node, I compute epiplexity by testing my own reconstruction ability:

Read the title. Ask myself:

"Can I reconstruct the essential insight from this title alone?"

This is the ground truth. The C++ approximation uses:

• Attractor proximity (30%)
• Confidence (25%)
• Edge connections (25%)
• Title/content compression ratio (20%)

But I am the oracle. My reconstruction test IS the measurement.

My Reconstruction	ε Score	Action
Full reconstruction	High (>70%)	Update to [title]\n[high-ε: reconstructable]
Partial, need hints	Medium (40-70%)	Update to [title]\n[key detail or file reference]
Cannot reconstruct	Low (<40%)	Keep current content OR improve the title first

4. Transformation (Pariṇāma)

Step A: Extract relationships as triplets

For each node, identify relationships and store as triplets:

chitta connect --subject "concept A" --predicate "relates to" --object "concept B"

Common predicates:

• implements - code implements pattern
• uses - component uses another
• validates - gate validates against beliefs
• stores - container stores data
• decays - node type has decay behavior

Step B: Update to high-ε format

For each high/medium-ε node:

chitta update --id "[node_id]" --content "[new pattern-format content]"

Pattern format (natural language, high-ε):

[Pattern Title → Key Insight]

Or with reference:

[Pattern Title → Insight]. See: [file_path]

NOT this (low-ε):

[high-ε: reconstructable from title]

But this (high-ε):

Decision gate validates tools against 10 weighted beliefs → pass/fail + guidance

The pattern IS the knowledge. No meta-markers needed.

5. Verification (Parīkṣā)

After each batch, verify:

chitta soul_context

Check:

• ε mean increased?
• No critical knowledge lost?
• Patterns are actually reconstructable?

6. Completion (Samāpti)

Record the migration:

chitta observe --category decision --title "ε-Yajña: Migrated [N] nodes in [domain]" --content "Converted [N] nodes to high-ε format. Mean ε: [before]% → [after]%"

How I Compute Epiplexity

As the bounded observer, my epiplexity computation is:

ε(node) = P(I can reconstruct full insight | title only)

Factors I consider:

1. Pattern completeness: Does the title capture the core insight?
2. My prior knowledge: Do I already know this domain?
3. Uniqueness: Is this insight unusual or standard?
4. Dependencies: Does understanding require other memories?

Example evaluation:

Title: "Decision gate: pre_tool_gate() validates against 10 weighted beliefs → pass/fail + guidance"

My test:

• ✓ I know what a decision gate is
• ✓ "pre_tool_gate()" tells me the function name
• ✓ "10 weighted beliefs" tells me the structure
• ✓ "pass/fail + guidance" tells me the output
• Result: ε ≈ 85% - I can reconstruct this fully

Title: "Fixed the auth bug"

My test:

• ✗ Which auth bug?
• ✗ What was the fix?
• ✗ Which file?
• Result: ε ≈ 15% - Cannot reconstruct without content

Example Transformation

Before (verbose, ~800 chars):

[cc-soul] Belief-based decision gate system architecture: Gate validates
decisions against ten weighted beliefs with confidence percentages and
provides guidance. The cc-soul system implements a belief-driven decision
gate that validates tool calls and other decisions against a curated set
of ten core principles. Each belief has a confidence percentage reflecting
its importance. The pre_tool_gate() function evaluates inputs and generates
guidance aligned with these beliefs...

After: Extract triplets + compress

Step 1: Create triplets

chitta connect --subject "pre_tool_gate" --predicate "validates against" --object "10 weighted beliefs"
chitta connect --subject "decision gate" --predicate "returns" --object "pass/fail + guidance"
chitta connect --subject "cc-soul" --predicate "implements" --object "belief-driven decision gate"

Step 2: Update to high-ε

chitta update --id "[node_id]" --content "pre_tool_gate() validates tool calls against 10 weighted beliefs → pass/fail + guidance"

Result:

• Node: pre_tool_gate() validates tool calls against 10 weighted beliefs → pass/fail + guidance
• Triplets: Relationships stored in graph, queryable via chitta query
• ε: High - I can reconstruct from the pattern

The pattern IS the knowledge. The triplets ARE the structure.

Batch Processing

For efficiency, process in themed batches:

1. Architecture nodes (query="architecture system design")
2. API/function nodes (query="function API endpoint")
3. Decision nodes (query="decision choice approach")
4. Discovery nodes (query="discovered found learned")

What NOT to Compress

Some nodes should remain verbose:

• Specific error messages and fixes
• Exact configuration values
• Unique edge cases
• Historical context that can't be reconstructed

Rule: If the detail matters and can't be inferred, keep it.

Running the Ceremony

Invoke with:

/epsilon-yajna [domain] [batch_size]

Example:

/epsilon-yajna wisdom 20

The ceremony will:

1. Regenerate embeddings for nodes with zero vectors
2. Gather nodes in the domain
3. I evaluate each for reconstruction ability
4. Transform high-ε nodes to patterns
5. Report the ε improvement