Skip to content

Architecture Diagram — Character Identity Protocol

This document provides visual representations of CIP’s three core mechanisms: Anchor Attractor, Cycle Stabilization, and Identity Gates.

All diagrams are operational abstractions. They do not represent proprietary model internals.


1. Anchor Attractor Model

How the anchor guides the model toward a stable identity state.

flowchart TD
    A[Training Distribution] --> B[Model Exploration]
    B --> C["Anchor Attraction
(reference image dominates)"]
    C --> D[Identity Convergence]
    D --> E{Identity Gates}
    E -->|PASS| F[Accepted Output]
    E -->|FAIL| G[Hard Abort → Re-binding]
    G --> B

Key insight:
The anchor does not bypass model optimization.
It introduces a previously validated solution state that attracts reconstruction toward a known stable region.
The model optimizes toward that prior rather than reconstructing freely.


2. Cycle Stabilization Model

How identity stability is maintained across multiple generation cycles.

flowchart LR
    A["Cycle A
Convergence Window"] --> B["Drift Accumulation
(stochastic sampling)"]
    B --> C["Context Stability
Threshold"]
    C --> D["Re-binding
(anchor re-injection)"]
    D --> E["Cycle B
Re-Converged Window"]
    E --> F["Drift Accumulation"]
    F --> G["Re-binding"]
    G --> H["Cycle C
Re-Converged Window"]

Key insight:
Stability is not permanent.
It is chained through disciplined re-convergence cycles.

[ Stable State A ]
        ↓
  Drift Accumulation
        ↓
  [ Re-Convergence ]
        ↓
[ Stable State B ]
        ↓
  Drift Accumulation
        ↓
  [ Re-Convergence ]
        ↓
[ Stable State C ]

3. Identity Gates — Operational Control Flow

How generation outputs are accepted or rejected.

flowchart TD
    A["Anchor Image + Minimal Prompt"] --> B[Generation Engine]
    B --> C[Output]
    C --> D{Face Gate}
    D -->|PASS| E{Skeleton Gate}
    D -->|FAIL| Z[Hard Abort]
    E -->|PASS| F{Proportion Gate}
    E -->|FAIL| Z
    F -->|PASS| G[Accepted Output]
    F -->|FAIL| Z
    Z --> H[Re-binding]
    H --> A

Gate policy:

PASS ⇔ Face Gate ∧ Skeleton Gate ∧ Proportion Gate

All gates must pass. Any single failure triggers Hard Abort.

The similarity threshold is operator-defined. The ≈90% value used in demonstrations refers to human-judged identity similarity relative to the anchor — not an automated metric or fixed protocol requirement.


4. Full CIP Operational Architecture

The complete control loop.

flowchart LR
    A["Anchor Image + Minimal Prompt"] --> B[Generation Engine]
    B --> C["Identity Gates
Face Gate
Skeleton Gate
Proportion Gate"]
    C -->|PASS| D[Accepted Output]
    C -->|FAIL| E[Hard Abort]
    E --> F[Re-binding]
    F --> G[Re-convergence]
    G --> B

5. Generation Pipeline — Without vs. With CIP

Without CIP:

User Prompt
      ↓
┌─────────────────────┐
│  Language Layer      │  Language interpretation
└─────────┬───────────┘
          ↓
┌─────────────────────┐
│  Reconstruction      │  ← Unconstrained
│  A → A'              │  Identity drift emerges here
└─────────┬───────────┘
          ↓
     Output (A')
     (Identity: uncontrolled)

With CIP:

Minimal Prompt  +  Anchor Image ──────────────┐
      ↓                                        │
┌─────────────────────┐                        │
│  Language Layer      │  Reduced load          │
│  (minimal prompt)    │  (fewer constraints)   │
└─────────┬───────────┘                        │
          ↓                                    ↓
┌─────────────────────────────────────────────┐
│  Reconstruction A → A'                       │
│  Anchor guides reconstruction                │  ← Operationally constrained
│  toward a previously converged solution state│
└─────────┬───────────────────────────────────┘
          ↓
     Output converges toward Anchor
          ↓
     Identity Gates (PASS / FAIL)
          ↓
   PASS → Production
   FAIL → Hard Abort → Re-binding

Clarification:
Language Layer / Reconstruction / Execution are conceptual abstractions for explanatory purposes. They do not imply knowledge of proprietary model internals.


6. Identity Drift Timeline

How identity similarity degrades over turns and how CIP intervenes.

flowchart TD
    T1["Turn 1
Identity: PASS ✓"] --> T2["Turn 2
Identity: PASS ✓"]
    T2 --> T3["Turn 3
Identity: PASS ✓"]
    T3 --> T4["Turn 4
Identity: Degrading"]
    T4 --> T5["Turn 5
Identity: FAIL ✗"]
    T5 --> HA[Hard Abort]
    HA --> RB[Re-binding
Anchor re-injected]
    RB --> C1["Cycle B — Turn 1
Identity: PASS ✓"]
    C1 --> C2["Cycle B — Turn 2
Identity: PASS ✓"]
    C2 --> C3["Cycle B — Turn 3
Identity: PASS ✓"]

Key insight:
Identity similarity is not maintained indefinitely.
It degrades gradually through stochastic drift.
CIP detects failure at the gate level and immediately re-binds the anchor.

High  ┤  ████ ████ ████ ░░░░
      │
      │                      FAIL
Low   ┤                        │
      │                        ↓
      │                  Hard Abort
      │                        ↓
High  ┤                  Re-binding
      │                        │
      │              ████ ████ ████ ░░░░
      └─────────────────────────────────→ Turns
           Cycle A             Cycle B

Operational Summary

Element Role
Anchor Image Primary convergence attractor
Minimal Prompt Auxiliary identity constraint
Identity Gates Operational validation (Face ∧ Skeleton ∧ Proportion)
Hard Abort Drift containment — immediate termination
Re-binding Anchor re-injection to restart convergence
Cycle Bounded convergence window

CIP does not control the model.
It controls the conditions under which the model converges.


See: Technical Mechanism for theoretical framing.
See: Quality Gate Addendum for gate definitions.

7. Persistent Anchor Layer (PAL) — Cross-Platform Architecture

PAL extends CIP’s inference-time anchoring into cross-session infrastructure.

PAL Layer Hierarchy

PAL (Persistent Anchor Layer)
  └── HDLA (High-Density Latent Anchoring)
        └── ARCM (Anchor Re-Convergence Method)

Platform Mapping

┌─────────────────────────────────────────────────────────────┐
│                  PAL — Cross-Platform View                  │
├──────────────┬──────────────────────┬───────────────────────┤
│   Platform   │    PAL Feature       │   CIP Layer           │
├──────────────┼──────────────────────┼───────────────────────┤
│  ChatGPT     │  Library (PDF/files) │  PAL Infrastructure   │
├──────────────┼──────────────────────┼───────────────────────┤
│  Claude      │  Project Files       │  PAL Infrastructure   │
├──────────────┼──────────────────────┼───────────────────────┤
│  Gemini      │  Context Caching     │  PAL Infrastructure   │
│              │  Gems / Ref Files    │  HDLA                 │
│              │  System Instructions │  ARCM Logic           │
└──────────────┴──────────────────────┴───────────────────────┘

Session Lifecycle with PAL

flowchart TD
    PAL["PAL — Persistent Anchor Layer
    (Library / Project Files / Context Cache)"]

    S1[Session Start]
    LOAD[Load Anchor + UID from PAL]
    GEN[Generate]
    GATE{Identity Gates}
    PASS[Accept Output]
    ABORT[Hard Abort]
    REBIND[Re-bind from PAL]
    END[Session End]

    PAL -->|provides anchor| S1
    S1 --> LOAD
    LOAD --> GEN
    GEN --> GATE
    GATE -->|PASS| PASS
    GATE -->|FAIL| ABORT
    ABORT --> REBIND
    REBIND -->|anchor restored from PAL| GEN
    PASS --> END

PAL eliminates full re-injection at session start and provides a stable anchor recovery source across Hard Abort cycles.

See: Column: Persistent Anchor Layer (PAL)