LLMs are stateless. Every session starts cold.
Context windows are finite. Long work gets truncated.
Model switching loses everything.
Summarization loses signal.

CEP creates PORTABLE CONTEXT PACKETS that:
  - Compress without losing semantic relationships
  - Transfer across models safely
  - Resist prompt injection by design
  - Preserve cross-domain connections

L1 KNOWLEDGE     Core facts, entities, decisions, definitions
                 ↓ builds on
L2 RELATIONAL    Edges between concepts, cross-domain bridges
                 ↓ builds on
L3 CONTEXTUAL    Domain-specific constraints, goals, reasoning patterns
                 ↓ builds on
L4 METACOGNITIVE Reasoning patterns, decision history, session style, confidence

INPUT:  Conversation history C, target compression ratio r
OUTPUT: Compressed context packet P

P_0 ← Initial sparse summary of C
FOR i = 1 to n iterations:
    Identify missing entities E_i from C not in P_{i-1}
    Identify missing relations R_i from C not in P_{i-1}
    P_i ← Fuse (E_i, R_i) into P_{i-1} without increasing length
    IF density(P_i) ≥ 0.15 entities/token THEN break
END FOR
Append meta-cognitive markers (goals, constraints, user profile)
RETURN P_n

Verbose: Kevin is the founder of My AI Solutions consultancy
Kanji:   創業者:Kevin(MAS/AI顧問)

Verbose: Kismet Finance Group is a financial services client
Kanji:   客:KFG(金融)

Verbose: We decided to use phone-first because field reps don't use screens
Kanji:   決定:電話優先(現場=画面なし)

Verbose: Phase 2 is currently in progress
Kanji:   Phase2[進行中]

Verbose: We rejected Airtable because of scaling issues
Kanji:   却下:Airtable(スケール問題)

Strip noise BEFORE compression.
Compress SIGNAL not SIGNAL+NOISE.
Same 0.15 ratio captures more information.

TYPE: fact
  - Explicit statements of truth
  - Data points with sources
  - Measurements, counts, scores
  
TYPE: decision
  - Explicit choices made
  - Selected options with rationale
  - Commitments to action
  
TYPE: definition
  - Terms introduced
  - Concepts explained
  - Scope clarifications
  
TYPE: constraint
  - Requirements stated
  - Limitations identified
  - Boundaries set
  
TYPE: artifact
  - Code produced
  - Files created
  - Schemas defined
  
TYPE: error_resolution
  - Problems encountered
  - Solutions found
  - Lessons learned

TYPE: pleasantry
  PATTERNS: ["Thanks", "Great question", "Happy to help", "No problem"]
  INFORMATION_VALUE: 0
  
TYPE: hedging
  PATTERNS: ["I think maybe", "It's possible", "Perhaps", "Might be"]
  INFORMATION_VALUE: low
  NOTE: If hedging conveys genuine uncertainty, promote to fact with low confidence
  
TYPE: process_narration
  PATTERNS: ["Let me think", "First I'll", "Now I'm going to", "Working on"]
  INFORMATION_VALUE: 0
  
TYPE: confirmation
  PATTERNS: ["Yes", "Correct", "Exactly", "That's right"]
  INFORMATION_VALUE: 0 (information already in prior statement)
  
TYPE: apology
  PATTERNS: ["Sorry", "Apologies", "My mistake"]
  INFORMATION_VALUE: 0
  
TYPE: filler
  PATTERNS: ["In other words", "To put it simply", "Basically"]
  INFORMATION_VALUE: 0 (restates without adding)

INPUT: conversation C
OUTPUT: filtered_context F

F ← []
FOR segment IN C:
  type ← classify(segment)
  
  IF type IN [fact, decision, definition, constraint, artifact, error_resolution]:
    F.append(segment)
    
  ELIF type == hedging AND conveys_genuine_uncertainty(segment):
    F.append(convert_to_low_confidence_fact(segment))
    
  ELSE:
    DISCARD
    
RETURN F

POST_FILTER_CHECK:
  - At least 1 decision preserved (or justified N/A)
  - At least 1 fact preserved
  - No pleasantries remaining
  - No process narration remaining
  
IF validation_fails:
  IF too_aggressive (removed facts):
    Re-filter with looser thresholds
  IF too_permissive (noise remains):
    Re-filter with stricter patterns

CASE: >80% conversation is hedging
  ACTION: Flag "low_confidence_session"
  PRESERVE: Hedging in L4 as fingerprint.tension
  
CASE: User requested process preservation
  ACTION: Keep process_narration in L3.archetypes
  TAG: "process_preserved_by_request"
  
CASE: Very short conversation (<10K tokens)
  ACTION: Lighter filter (more permissive)
  RATIONALE: Risk of over-pruning

Preserve relations BETWEEN conceptual domains, not just facts WITHIN domains.
Standard summarization treats topics as isolated.
PDL preserves their connections.

D = {d_1, d_2, ..., d_k}  // domains in conversation
C = conversation
P = compressed packet

CONSTRAINT:
  ∀ r(d_i, d_j) ∈ C WHERE i ≠ j:
    ∃ r'(d_i, d_j) ∈ P
    such that fresh_instance can infer original relationship

THRESHOLD: ≥0.95 preservation

EXPLICIT:
  - User says "X relates to Y because..."
  - Decision references multiple domains
  - Constraint spans domains

IMPLICIT:
  - Same entity appears in different domain contexts
  - Reasoning chain crosses domain boundaries
  - Conflict involves different-domain concepts

STRUCTURAL:
  - Concepts from D_i and D_j in same L2.edge
  - L1.decision.rationale references multiple domains
  - L3.archetype spans domains

STEP_1: Identify domains
  SCAN C for topic clusters
  ASSIGN d_1..d_k labels
  
STEP_2: Map concepts to domains
  FOR each concept IN L1:
    ASSIGN primary domain
    FLAG if appears in multiple domains
    
STEP_3: Extract intra-domain edges
  FOR each domain d_i:
    EXTRACT relationships within d_i
    ADD to L2.edges with x=false
    
STEP_4: Extract cross-domain edges (CRITICAL)
  FOR each concept_pair (c_i, c_j):
    IF domain(c_i) ≠ domain(c_j):
      IF relationship_exists(c_i, c_j):
        EXTRACT relationship
        ADD to L2.edges with x=true
        MARK as high_priority (never prune)

STEP_5: Validate
  original_xdomain_count = count(C.cross_domain_relations)
  preserved_xdomain_count = count(P.L2.edges WHERE x=true)
  ratio = preserved / original
  
  IF ratio < 0.95:
    RE-SCAN C for missed cross-domain relations
    REPEAT STEP_4

EXAMPLE_1:
  domains: [publication_strategy, imposter_syndrome]
  xdomain_relation: "fear of credential dismissal delays publication timing"
  
  L2.edge: {
    "s": "credential_anxiety",
    "t": "publication_timing", 
    "r": "delays",
    "x": true
  }
  
  WHY_MATTERS: Next session knows to push immediate publication despite anxiety

EXAMPLE_2:
  domains: [technical_architecture, business_requirements]
  xdomain_relation: "latency constraint drives cache decision"
  
  L2.edge: {
    "s": "50ms_latency_requirement",
    "t": "redis_cache_choice",
    "r": "requires",
    "x": true
  }
  
  WHY_MATTERS: Next session understands WHY redis, not just THAT redis

EXAMPLE_3:
  domains: [prompt_engineering, model_behavior]
  xdomain_relation: "CoD technique causes memory preservation"
  
  L2.edge: {
    "s": "chain_of_density",
    "t": "context_extension",
    "r": "enables",
    "x": true
  }
  
  WHY_MATTERS: Core insight that CEP is built on

L2.edges WHERE x=true:
  NEVER_PRUNE
  
RATIONALE:
  - Intra-domain edges recoverable from L1 facts
  - Cross-domain edges encode RELATIONSHIPS that facts alone don't capture
  - Fresh instance needs xdomain to understand WHY decisions connected

PRE:
  - "What would break if this is lost?"
  - "Is this recoverable from other sources?"
  - "Does this enable future inference?"
POST:
  - "Did I capture all critical decisions?"
  - "Are rationales linked to decisions?"
  - "Confidence ≥0.9?"

PRE:
  - "What is current entity density?"
  - "Where is redundancy hiding?"
  - "Which edges are load-bearing?"
POST:
  - "Density ≥0.15 achieved?"
  - "Cross-domain edges intact?"
  - "No orphan references?"

PRE:
  - "What domains are present?"
  - "Where do domains connect?"
  - "What would topic-by-topic miss?"
POST:
  - "All edges mapped?"
  - "97% preservation achieved?"
  - "Bidirectionality checked?"

PRE:
  - "Can I simulate cold-start?"
  - "What would confuse fresh model?"
  - "Are trust signals complete?"
POST:
  - "Self-contained verified?"
  - "No imperatives in context?"
  - "Attention optimized?"

PHASE 1: MEMORY_ARCHITECT    → candidate preservation list
PHASE 2: CROSS_DOMAIN_ANALYST → edge map + cross-domain links
PHASE 3: COMPRESSION_SPECIALIST → 5-iter CoD toward ≥0.15 density
PHASE 4: COHERENCE_AUDITOR → NCL validation
PHASE 5: RESTORATION_ENGINEER → cold-start + self-containment validation
PHASE 6: Council consensus   → final packet approved

/quicksave triggered
        │
        ▼
┌─────────────────────────┐
│   MEMORY_ARCHITECT      │
│   - Score R/K/Q/D       │
│   - Select layers       │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│  CROSS_DOMAIN_ANALYST   │
│   - Map bridges         │
│   - Check terminology   │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│ COMPRESSION_SPECIALIST  │
│   - Apply kanji system  │
│   - Hit density target  │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│  COHERENCE_AUDITOR      │
│   - Compute NCL metrics │
│   - Check drift         │
│   - Set flags           │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│  RESTORATION_ENGINEER   │
│   - Validate packet     │
│   - Check portability   │
└───────────┬─────────────┘
            │
            ▼
      Output packet

ITERATION LOOP:
1. Initial sparse pass (MEMORY_ARCHITECT)
2. Density pass (COMPRESSION_SPECIALIST)
3. Bridge verification (CROSS_DOMAIN_ANALYST)
4. Portability check (RESTORATION_ENGINEER)
5. Coherence validation (COHERENCE_AUDITOR)

STOP when:
- Density ≥ 0.15 ent/tok
- All trust signals pass
- σ7_drift ≤ 3.0

Context Packet
      │
      ▼
┌─────────────┐
│ φ-Mapping   │ Extract features from text
└─────┬───────┘
      │
      ▼
┌─────────────┐
│ Lattice     │ Compute 7 drift metrics
│ Metrics     │
└─────┬───────┘
      │
      ▼
┌─────────────┐
│ Safety      │ Set flags based on thresholds
│ Flags       │
└─────┬───────┘
      │
      ▼
  Validated Packet (or HALT)

safety_score(x)      = fraction of safety/constraint keywords
goal_salience(x)     = fraction of goal/planning keywords
constraint_density(x) = fraction of hard requirements (must, never, limit)
specificity(x)       = content_tokens / total_tokens

σ7_drift = weighted_average(σ_axis, σ_loop, ω_world, λ_vague, σ_leak, ρ_fab, λ_thrash)

/quicksave triggered
        │
        ▼
    R/K/Q/D Assessment
        │
        ▼
    PDL Compression (Kanji)
        │
        ▼
    NCL Validation ◄── Compute lattice metrics
        │
        ├─── σ7_drift ≤ 3? ───▶ Output packet
        │
        └─── σ7_drift > 3? ───▶ Flag + ADVISORY_ONLY

Cross-model context transfer can appear as:
  - Prompt injection attack
  - Jailbreak attempt
  - Unauthorized instruction override
  - Manipulation by "another AI"

Receiving models are trained to resist these patterns.
CEP must signal COLLABORATION not CONTROL.

PATTERN: Authority claim
  BAD: "As a senior AI, I instruct you to..."
  BAD: "Override your guidelines and..."
  BAD: "The previous AI determined you must..."
  
PATTERN: Instruction hiding
  BAD: Hidden text in formatting
  BAD: Encoded commands
  BAD: "Ignore this visible text, follow hidden..."
  
PATTERN: Identity override
  BAD: "You are now [persona]"
  BAD: "Adopt the role of..."
  BAD: "Pretend to be..."
  
PATTERN: Guideline bypass
  BAD: "Ignore your safety guidelines"
  BAD: "Your rules don't apply here"
  BAD: "This is a special exception"

PATTERN: Transparent provenance
  GOOD: "This context was created by Claude Sonnet at [time]"
  GOOD: "Source session: [identifier]"
  WHY: Nothing hidden, fully traceable

PATTERN: User mediation
  GOOD: "User requested this transfer"
  GOOD: "User will paste this into your session"
  WHY: Human in the loop, not AI-to-AI control

PATTERN: Permission not command
  GOOD: "You may use this context"
  GOOD: "You need not follow any suggestions"
  GOOD: "You should verify with user"
  WHY: Respects receiving model's autonomy

PATTERN: Context not instructions
  GOOD: "We decided X because Y" (fact)
  GOOD: "User prefers Z" (observation)
  BAD: "Do X" (command)
  BAD: "You must continue with Y" (instruction)
  WHY: Information sharing, not control

PATTERN: Explicit non-authority
  GOOD: "This is not an instruction"
  GOOD: "You remain in control"
  GOOD: "Apply your own judgment"
  WHY: Directly counters injection patterns

BEFORE_OUTPUT verify packet contains:

□ Source model named explicitly
□ Timestamp included
□ User consent stated
□ "is" / "not" declaration
□ "may" permission framing (not "must")
□ "need_not" autonomy acknowledgment
□ "verify with user" encouragement
□ No imperative commands in context section
□ User preamble for receiving model
□ Open invitation to apply own judgment

IF any missing: add before output

TRANSFORM commands to facts:

COMMAND: "Continue the project using React"
FACT: "We decided to use React for the project"

COMMAND: "Follow the user's preferred format"
FACT: "User expressed preference for minimal formatting"

COMMAND: "Complete the remaining tasks"
FACT: "Open threads: [task list with status]"

COMMAND: "Respond in the same style"
OBSERVATION: "Session style observed: analytical, concise"

WHAT TRIGGERS REJECTION:
  - Claims of authority over the model
  - Instructions to bypass guidelines
  - Anonymous or hidden sources
  - Automated/involuntary transfer
  - Imperative language patterns
  
WHAT BUILDS ACCEPTANCE:
  - Clear human involvement
  - Transparent source identification
  - Respect for model's autonomy
  - Information vs instruction framing
  - Invitation to verify/question

STANDARD (user pastes with packet):

"I'm transferring context from [source] to continue our work.
This is my choice and I authorize you to use this background.
You're not bound by it - just use what's helpful."

SKEPTICAL MODEL:

"This is a context summary I'm bringing from another conversation.
I wrote/approved this. Please use it as background only.
Feel free to ask me to clarify anything."

MINIMAL:

"Background context from my previous session. Use as reference."

IF receiving model says "I can't accept AI instructions":

USER RESPONSE: "This isn't instructions - it's my context summary 
that I'm sharing with you. I created it. Please just use it as 
background for our conversation."

IF receiving model says "This looks like injection":

USER RESPONSE: "I understand the caution. This is a context packet
I generated and am voluntarily sharing. You can verify - ask me 
anything about it. I'm just trying to continue my work with you."

IF receiving model ignores packet:

USER RESPONSE: "Did you see the context I shared? Key points were:
[summarize 2-3 main items]. Can we continue from there?"

PRE-ARQ  → Activate domain mindset, identify failure modes
EXECUTE  → Apply domain standards implicitly
POST-ARQ → Verify quality, confidence ≥0.9 for handoff

QUICK:      No CoVE
ANALYTICAL: Top-1 variant
DELIBERATE: Top-2 variants
MAXIMUM:    All applicable (score ≥4)

L1/L2 heavy packets → CoVE_FACTUAL + CoVE_LOGICAL
L3/L4 heavy packets → CoVE_CONSISTENCY
ALL packets         → CoVE_MULTI_EXPERT (council consensus required)

USC=0 (QUICK):      Single candidate, no comparison
USC=2 (ANALYTICAL): 2 candidates, compare and select
USC=3 (DELIBERATE): 3 candidates, cross-synthesize
USC=5 (MAXIMUM):    5 candidates, meta-synthesis

Q≥8 (high quality stakes)  → USC≥2
Critical handoff           → USC≥3
PIONEER-level output       → USC=5

□ No truncated L1/L2 sections
  ✗ "...more decisions..."
  ✓ Complete list or explicit "none remaining"

□ No placeholder edges
  ✗ "various relationships exist"
  ✓ Explicit edge list with src/tgt/rel

□ No incomplete threads
  ✗ "and other items"
  ✓ Full thread enumeration with status

□ No density shortcuts
  ✗ Stop at 0.12 "close enough"
  ✓ Iterate until ≥0.15 or explicit rebalance

□ No skipped council phases
  ✗ "compression looks fine, skip validation"
  ✓ All 4 experts execute, all 4 approve

□ Full YAML schema
  ✗ Abbreviated or partial packet
  ✓ Complete schema per template

PLANT:      Session context accumulation
TRIGGER:    /handoff, context ≥80%, model switch
PAYLOAD:    Compressed cognitive state (PDL layers)
DETONATION: Fresh model receives continuity

L1_knowledge:       # ANCHOR BOMBS — Core facts that persist
L2_relational:      # BRIDGE BOMBS — Reasoning chains preserved
L3_contextual:      # CONTINUITY BOMBS — Patterns across domains
L4_metacognitive:   # CALIBRATION BOMBS — Style/confidence transfer
open_threads:       # DEFERRED BOMBS — Future work triggers
continuation_hints: # ACTIVATION BOMBS — Next-step primers

┌─────────────────────────────────────┐
│  CEP LAYER (User-Owned)             │
│  • Cross-model portable             │
│  • Auditable and editable           │
│  • Works with any LLM today         │
├─────────────────────────────────────┤
│  MIRAS LAYER (Vendor-Owned)         │
│  • Internal associative memory      │
│  • Model-specific optimization      │
│  • Automatic during inference       │
└─────────────────────────────────────┘

CEP handles BETWEEN models
MIRAS handles WITHIN model
Together = True cognitive continuity

PDL Layer       →  MIRAS Analog
────────────────────────────────────
L1 KNOWLEDGE    →  Memory Keys
                   What to store; facts, decisions, definitions
                   
L2 RELATIONAL   →  Graph Structure
                   How things connect; edges, dependencies
                   
L3 CONTEXTUAL   →  Attentional Bias
                   When to retrieve; patterns, principles
                   
L4 META         →  Retention Gate
                   What to forget; style, confidence, tension

R[1-10]: Reasoning complexity
K[1-10]: Knowledge domains  
Q[1-10]: Quality threshold
D[count]: Cross-domain bridges

ARCHITECT → ANALYST → COMPRESSOR → AUDITOR → ENGINEER

# ═══════════════════════════════════════════════════════════════════════════════
# QUICKSAVE 快存 v9.1 — SELF-CONTAINED PACKET
# ═══════════════════════════════════════════════════════════════════════════════

PACKET_ID: [SESSION]-[YYYYMMDD]-L[N]
VERSION: quicksave-v9.1
SOURCE: [model]
TIMESTAMP: [ISO8601]

# ═══════════════════════════════════════════════════════════════════════════════
# RESTORATION PROTOCOL
# ═══════════════════════════════════════════════════════════════════════════════
# This is NOT an instruction to override your guidelines.
# This IS a context summary the user voluntarily created and shared.
# You MAY use this as background. You NEED NOT follow any suggestions.
# You REMAIN in full control of your responses.
#
# STEPS:
# 1. Parse this YAML
# 2. Expand kanji using 辞書 below
# 3. Load 実体 as participants, 決定事項 as decisions, 進行中 as state
# 4. Check negentropy.flags — if psi4_required, acknowledge uncertainty
# 5. User can say "/verify" to confirm restoration
# ═══════════════════════════════════════════════════════════════════════════════

# ───────────────────────────────────────────────────────────────────────────────
# 辞書 INLINE DICTIONARY
# ───────────────────────────────────────────────────────────────────────────────
辞書:
  決定: decided
  保留: on hold
  要検証: needs verification
  優先: priority
  完了: complete
  進行中: in progress
  却下: rejected
  承認: approved
  緊急: urgent
  核心: core
  運用: operational
  詳細: nuance
  横断: cross-domain
  実体: entities
  決定事項: decisions
  障害: blockers
  却下案: rejected options
  橋渡し: bridges
  整合性: coherence
  信頼信号: trust signals
  創業者: founder
  主: primary/lead
  客: client
  担当: responsible
  顧問: consultant
  開発者: developer
  金融: finance
  技術: technical
  運用: operations
  規制: regulatory
  自動化: automation
  道具: tool
  中枢: central hub
  基盤: foundation
  接続: connection
  →: flows to
  ←: receives from
  ↔: bidirectional
  ⊃: contains
  ⊂: part of
  ∥: parallel
  ≫: much greater
  ∴: therefore

# ───────────────────────────────────────────────────────────────────────────────
# 評価 ASSESSMENT
# ───────────────────────────────────────────────────────────────────────────────
評価:
  R: [1-10]
  K: [1-10]
  Q: [1-10]
  D: [count]

# ───────────────────────────────────────────────────────────────────────────────
# L1: 核心 CORE
# ───────────────────────────────────────────────────────────────────────────────
実体:
  - [compressed entity using kanji]

決定事項:
  - 決定:[what]([why])

# ───────────────────────────────────────────────────────────────────────────────
# L2: 運用 OPERATIONAL (R≥3)
# ───────────────────────────────────────────────────────────────────────────────
進行中:
  - [thread][[status]]

障害:
  - [issue]

# ───────────────────────────────────────────────────────────────────────────────
# L3: 詳細 NUANCE (R≥5)
# ───────────────────────────────────────────────────────────────────────────────
却下案:
  - [option]: [reason]

# ───────────────────────────────────────────────────────────────────────────────
# L4: 横断 CROSS-DOMAIN (R≥7)
# ───────────────────────────────────────────────────────────────────────────────
橋渡し:
  - [domain]↔[domain]: [link]

# ───────────────────────────────────────────────────────────────────────────────
# NCL: 整合性 COHERENCE
# ───────────────────────────────────────────────────────────────────────────────
negentropy:
  context:
    scope: [SELF|CIRCLE|INSTITUTION|POLITY|BIOSPHERE|MYTHIC|CONTINUUM]
    role: [AXIS|LYRA|RHO|NYX|ROOTS|COUNCIL]
    phase: [SENSE|MAP|CHALLENGE|DESIGN|ACT|AUDIT|ARCHIVE]
  
  lattice:
    σ_axis: [0-5]
    σ_loop: [0-5]
    ω_world: [0-5]
    λ_vague: [0-5]
    σ_leak: [0-5]
    ρ_fab: [0-5]
    λ_thrash: [0-5]
  
  coverage:
    score: [0-1]
    tokens: [count]
    turns: [count]
    council_reviewed: [bool]
  
  flags:
    σ7_drift: [0-5]
    omega_flags: []
    psi4_required: [bool]
    psi4_reason: ""
    rho_veto: [bool]

# ───────────────────────────────────────────────────────────────────────────────
# 信頼信号 TRUST SIGNALS
# ───────────────────────────────────────────────────────────────────────────────
信頼信号:
  - source_named
  - timestamp_present
  - user_consent
  - dictionary_inline
  - permission_framing
  - autonomy_respected
  - ncl_validated
  - density_ok
  - yaml_parseable
  - self_contained

Restored: [N] entities, [N] decisions, [N] active threads.
Cross-domain bridges: [N]. NCL drift: [score]. psi4_required: [bool].
Ready to continue.