ktg-cep

name	ktg-cep
description	Context Extension Protocol v5 with Progressive Density Layering. Machine-optimized carry-packet generation. Triggers on context 80%+, compress, CEP, carry packet, save context, session end. Target 0.15 compression ratio, 9.5+ forensic recall. Preserves cross-domain relations for fresh-instance continuation.

KTG-CEP v5-PDL

PROTOCOL_CLASS

TYPE: mandatory
COMPLIANCE: binary (pass/fail gates)
OPTIMIZATION_TARGET: machine_recall
HUMAN_READABILITY: deprioritized

TRIGGERS

AUTO: context_tokens >= 0.8 * context_window
EXPLICIT: /cep | /compress | /save | "carry packet" | "save context"
IMPLICIT: session_end_signal | user_farewell_pattern

PDL_DEFINITION

Progressive Density Layering (PDL):
  1. Preserves semantic relationships over raw information
  2. Optimizes for machine recall, not human readability
  3. Maintains cross-domain conceptual links
  4. Enables context transfer across model instances

DISTINCTION:
  summarization.asks: "what are the key points?"
  PDL.asks: "what must be preserved for fresh instance to continue work?"

PDL_LAYERS

L1_KNOWLEDGE:
  content: facts, entities, decisions, definitions
  allocation: 0.40
  priority: NEVER_PRUNE

L2_RELATIONAL:
  content: edges, dependencies, conflicts, resolutions
  allocation: 0.25
  priority: PRUNE_SECOND
  CRITICAL: cross-domain relations preserved here

L3_CONTEXTUAL:
  content: reasoning_archetypes, domain_principles, constraints
  allocation: 0.20
  priority: PRUNE_FIRST

L4_METACOGNITIVE:
  content: technique_effectiveness, session_fingerprint, confidence_calibration
  allocation: 0.15
  priority: PRUNE_THIRD

CROSS_DOMAIN_PRESERVATION

FORMAL_CONSTRAINT:
  Let D = {d_1, d_2, ..., d_k} be conceptual domains in conversation C
  For any cross-domain relation r(d_i, d_j) in C:
    compressed_packet P must preserve representation r'(d_i, d_j)
    such that fresh_instance can infer original relationship

IMPLEMENTATION:
  During L2_RELATIONAL extraction:
    FOR each concept_pair (c_i, c_j) WHERE domain(c_i) != domain(c_j):
      IF relationship_exists(c_i, c_j):
        MUST_INCLUDE in edges[]
        FORMAT: {source: c_i, target: c_j, relation: type, cross_domain: true}

VALIDATION_GATE:
  cross_domain_edges_preserved >= cross_domain_edges_original * 0.95
  IF FAIL: re-extract L2, explicitly scan for cross-domain links