Glossary: Character Identity Protocol

Core terms used throughout the documentation.

Archetype Drift

A character drift phenomenon in which a generated character retains partial visual continuity with the anchor identity while shifting toward a stronger nearby archetype in the model’s reconstruction space. Archetype Drift is distinct from facial collapse or structural degradation. It is a directional failure: the character’s visual features remain substantially intact, but the identity interpretation consistency — the social register, personality impression, and role identity — shifts toward a more dominant nearby template. The result is captured by the phrase same face, different person.

As an example: a character defined as a reserved, precise intellectual may drift across generation cycles toward a warmer, more approachable register — not through any single dramatic change, but through gradual reweighting of expressive and social dimensions. The face remains recognizable. The person does not.

Archetype Drift is particularly significant in multi-scene, multi-session, and video generation workflows, where cumulative drift across outputs can produce a character who is visually continuous but identity-discontinuous. Standard structural gates — Face Gate, Skeleton Gate, Proportion Gate — may not detect this failure mode, as they evaluate feature similarity rather than identity register.

Detection of Archetype Drift may require evaluation beyond structural feature comparison, including assessment of social impression, expressive register, and role identity consistency against the anchor UID definition.

See: Identity Drift Taxonomy — Archetype Drift — PAL Hypothesis Document

Anchor-Sufficient Convergence (ASC)

Same UID. Different prompt. Same character.

A convergence condition in which character identity is maintained across generations using only a PAL-registered anchor asset and a stable UID — without requiring ControlNet, OpenPose, seed control, LoRA, or model fine-tuning.

Formal conditions:

Condition	Status
Same UID	Required
PAL-registered anchor	Required
ControlNet	Not required
OpenPose	Not required
Seed control	Not required
LoRA	Not required
Model fine-tuning	Not required

Implication:

ASC demonstrates that identity is not a property of generation tools — it is a property of the anchor.

When PAL is properly configured, the anchor asset alone is sufficient to maintain character identity across sessions, platforms, and generation cycles.

The anchor is sufficient. The tools are optional.

Related:

Anchor

Anchor = Minimal Prompt + Previously Converged Output Image

A previously generated image that represents a known, high-quality solution state.
Used as a constraint to bias future generation toward the same convergence point.

The image is not an artistic reference or stylistic inspiration. It is a record of a solution the model has already found.

Anchor Mechanism

The protocol-level stabilizing principle by which a validated anchor image is introduced as a high-information conditioning signal to bias reconstruction toward a previously achieved identity state.

Distinct from image-to-image (which encourages variation) and seed fixing (which primarily controls initial noise sampling).

Anchor Convergence

The operational workflow for forming and re-establishing an anchor. Includes: high-density sample selection, identity gate validation, identifier binding, minimal prompt reduction, and multi-view expansion. Constitutes the entry condition of the CIP governance loop.

Distinct from Anchor Mechanism: Anchor Mechanism is the stabilizing principle; Anchor Convergence is the formation and re-convergence workflow.

Infrastructure Condition

Within the PAL framework, an infrastructure condition refers to the deployment state in which anchor materials remain available at inference time through persistent file, project, caching, or reference systems.

Infrastructure condition describes the implementation layer of PAL — how anchor materials are made available.

Distinct from stabilization condition, which describes the functional effect of that availability.

PAL is an infrastructure condition. Its functional effect is a stabilization condition.

See: Stabilization Condition See: PAL Hypothesis Document

Identity Recall Probability

An operational likelihood that subsequent generation cycles converge toward the same identity state as the anchor. Increased by identifier binding, multi-view expansion, and minimal prompt discipline. Not a mathematical probability but a qualitative indicator of reconstruction stability.

Convergence / Convergence Point

The state at which a generative model’s output stabilizes into a coherent, consistent result.

A “convergence point” is a location in the model’s solution space where identity, structure, and rendering align with unusual precision.

The protocol does not force convergence — it biases the search space toward a previously achieved convergence point.

Cultural Frame Drift

A form of identity drift in which a generated output migrates toward the dominant cultural or aesthetic conventions of the model’s training distribution, away from the culturally bounded definition of the anchor identity.

Cultural Frame Drift is particularly observable when a character defined within a minority visual culture — such as a Japanese anime aesthetic — is generated in contexts where the training distribution is heavily weighted toward a different cultural convention.

The drift does not necessarily produce structural identity failure as detected by Face Gate, Skeleton Gate, or Proportion Gate. It may instead manifest as a shift in facial impression, expressive register, or stylistic rendering that moves the output away from its culturally defined anchor.

PAL may suppress Cultural Frame Drift by keeping the culturally bounded anchor definition persistently available at inference time.

See: PAL Hypothesis — Section 10

Convergence Recoverability

The ability of a character identity to be repeatedly reconstructed as the same identity under variation, reinterpretation, and context change — without requiring strong external stabilization.

Convergence Recoverability is distinct from:

Match rate — a single-generation similarity score
Identity gate PASS — a binary validation outcome
LoRA consistency — constraint-enforced similarity

It describes a character-level property of the identity itself, not of the control mechanism applied to it.

Characters with high Convergence Recoverability tend to share:

sufficient distinctiveness (clear identity markers)
without reconstruction overload (not pushed into sparse regions)
style-identity separation (style does not compete with identity)

Related:

Reconstruction Durability (RD)

A character-level property describing the likelihood that identity can be repeatedly recovered across generation cycles without external stabilization, determined by the character’s position within the model’s high-density reconstruction regions.

Reconstruction Durability is the combined property of:

high Convergence Recoverability
positioning in high-density reconstruction regions
style-identity separation

Key distinction: Reconstruction Durability is not a property of the control method. It is a property of how the character was formed.

Design implication:

Character Formation (RD)
  └── CIP Governance
        └── PAL Persistence
              └── ASC Conditions

A character designed for high Reconstruction Durability requires less governance overhead, produces more stable anchors, and exhibits higher Convergence Recoverability across platforms and model versions.

Related:

Reconstruction (A→A’)

The internal process by which a generative model interprets a prompt and reformulates it into a generation target.

User prompt (A) → Model's internal reformulation (A') → Generation

This stage is:

Invisible to the user
Not directly controllable
The primary source of identity drift

The anchor mechanism constrains A→A’ by supplying a high-information prior.

Layer A / Layer B / Layer C

A theoretical abstraction of the generation process — not a claim about proprietary architecture.

Layer	Function
A	Language interpretation
B	Reconstruction / optimization
C	Execution (latent sampling and rendering)

Verbose prompts activate Layers A and B more strongly.
Minimal prompts reduce pressure on Layers A and B, allowing the anchor to dominate.

Minimal Prompt

A prompt containing only factual, identity-defining attributes — no adjectives, no mood descriptors, no compositional instructions.

Purpose: to reduce optimization pressure from the model’s training priors, preserving the convergence state established by the anchor.

Example:

Japanese woman, black hair, brown eyes, 168cm, slender

Not:

beautiful, elegant, mysterious Japanese woman with flowing black hair...

Library Contamination

A condition in which anchor materials registered in a persistent reference layer contain content that produces harmful, misleading, or unintended reconstruction behavior.

Because PAL may function as a normative and cultural frame persistence layer, contaminated anchor materials may stabilize harmful normative frames rather than intended ones.

Library Contamination requires active purge of affected materials. Passive removal or deprecation is insufficient, as accessible materials may continue to influence reconstruction behavior.

The operator is responsible for anchor library governance, contamination detection, and purge procedures.

PAL does not validate anchor content. The operator does.

See: PAL Hypothesis — Section 10.7
See: ai-identity-governance

Match Rate

A human-assessed similarity score between the anchor image and a generated output.

Match rate is not an identity verification metric. It is an early warning indicator for character drift — a signal that degradation may be beginning before full collapse occurs.

Evaluated across three dimensions:

FaceGate — facial identity consistency
SkeletonGate — skeletal proportion and alignment
ProportionGate — overall body proportion

The similarity threshold is operator-defined. The ~90% figure is a demonstration value only — it is not a protocol standard. Treating it as a fixed threshold is a misapplication of this protocol.

Match rate measures the risk of drift, not the fact of identity.

Match rate is a human judgment, not an automated metric.

Normative Drift

A form of drift in which an AI system’s behavioral rules, decision priorities, or role definition shift across sessions — away from the anchored normative frame and toward the model’s statistical defaults.

Normative Drift is not limited to image generation. It may occur in conversational AI, agentic systems, and any AI deployment where behavioral consistency across sessions is operationally required.

Unlike structural identity drift, Normative Drift may not be detectable through visual gate evaluation. It requires assessment of behavioral posture, role consistency, and interactional register across outputs.

PAL may suppress Normative Drift by keeping a structured role and behavioral anchor persistently available at inference time.

See: PAL Hypothesis — Section 10
See: ai-identity-governance

Persistently Available

Within the CIP and PAL frameworks, “persistently available” describes anchor materials that remain accessible across session boundaries through platform-level persistence features.

What this means operationally:

The materials remain available at the start of a new session without explicit re-injection by the operator
Access is provided through persistent file, project, caching, or reference systems
Availability persists across session resets

What this does not mean:

Continuously loaded into active memory
Technically permanent or immutable storage
Parameter-level encoding within the model

The term refers to operational availability across sessions, not to any specific technical implementation of persistence.

See: Persistent Anchor Layer (PAL)

Quality Gate

The set of validation conditions that must all pass before a generation is accepted.

PASS = FaceGate ∧ SkeletonGate ∧ ProportionGate

If any gate fails → hard abort.
No progressive correction permitted.

See: Quality Gate Addendum

Hard Abort

The mandatory stop condition triggered when any quality gate fails.

The session is terminated. Contaminated outputs are discarded.
No continuation, no correction, no exception.

Contamination

The accumulation of identity drift in a session’s context, caused by:

Failed generations that were not discarded
Continued generation after drift detection
Excessive session length without re-anchoring

Contaminated sessions cannot be recovered by prompt adjustment.
They must be abandoned and restarted with a clean anchor.

UID (Unique Identifier)

A stable name or label assigned to a character’s converged state.

Purpose: enables cross-session recall without re-providing the full anchor each time.
Example: “Model Hana”, “Model A”, “Shizuka”

The UID corresponds to a specific converged visual state — not assigned arbitrarily.

Identity Drift

The gradual deviation of generated outputs from the established character identity across turns.

Caused by:

Reconstruction variability (A→A’) accumulating across turns
Optimization pressure from verbose prompts
Session context degradation
Insufficient anchoring frequency

Stabilization Condition

Within the PAL framework, a stabilization condition refers to the functional effect produced when anchor materials are persistently available at inference time.

Stabilization condition describes what PAL does — it may improve reconstruction consistency across sessions by maintaining anchor availability.

Distinct from infrastructure condition, which describes how PAL is implemented.

Usage guidance:

Use “stabilization condition” when describing the functional effect of PAL
Use “infrastructure condition” when describing the implementation layer of PAL

PAL is framed as an inference-time stabilization condition, not a training-time intervention.

See: Infrastructure Condition See: PAL Hypothesis Document

Style Drift

The deviation of generated outputs from the defined rendering regime — line weight, color quantization, shading abstraction, texture ceiling — toward a higher information density state (typically: anime abstraction → semi-realistic → photorealistic).

Style Drift is technically distinct from structural Identity Drift.

In general production contexts, style drift may be recoverable through re-anchoring in a subsequent cycle.

In style-defined identity domains — anime, manga, illustration, game, and franchise animation — style drift is classified as a failure at the point of occurrence, regardless of recoverability.

This classification rests on the industry standard that:

Rendering regime is constitutive of identity, not decorative
A recoverable failure is still a failure
Zero-tolerance enforcement applies to rendering regime deviation in IP-governed production

Style Drift in anime, manga, game, and serialized IP contexts is not aesthetic variation.
It is identity violation — and triggers Hard Abort under the same conditions as structural drift.

See: Video Addendum — Style-Layer Primacy Problem

Cross-Platform Migration

The procedure for re-establishing a character identity on a different generative platform.

The anchor image functions as a platform-agnostic identity record.
The character’s convergence state can be recovered on any system capable of image-guided generation.

See: Case 04: Cross-Platform Migration — “Shizuka”