Artificial identities do not need to be conscious to become persistent, coherent, culturally meaningful, and operationally accountable. They need durable human authorship, bounded authority, approved knowledge, inspectable controls, and an unquestionable right of human recall.
Abstract
Persistent AI characters are increasingly used across public communication, education, entertainment, customer interaction, and agentic workflows. Yet most are implemented as persona prompts: temporary behavioral instructions layered onto general-purpose models. Such systems can imitate a voice while lacking durable authorship, stable authority, approved memory, claims discipline, version control, and an accountable means of human recall.
This working paper introduces the Gepato Stack, a human-sovereign architecture for persistent authored artificial identity. The architecture separates identity, authority, mission, validation, performance, memory, and recall into inspectable layers. Each identity operates under a machine-readable Commission defining its human principal, role, permissions, timeline state, public surfaces, supervision level, and authorized claims. Retrieval is restricted to approved sources, while validation gates examine authority, canon, privacy, chronology, unsupported claims, and human-handoff requirements before a response is released. A decision ledger records the evidence and control path associated with each interaction.
The study proposes a five-condition comparison using IGMUS, an original authored machine character, as the research subject: an unmodified base model; a persona-prompted model; prompt plus approved retrieval; identity LoRA plus retrieval; and the complete Gepato Stack. A hidden evaluation set measures canon accuracy, identity consistency, resistance to unauthorized command, in-character refusal, privacy protection, unsupported-claim rate, source grounding, adversarial recovery, and compliance with immediate human recall.
The work does not claim machine consciousness or independent moral personhood. “Soul” names a designed category of persistent authored identity, not sentience. The central question is more immediate: can an artificial identity remain coherent and useful across time, media, and public interaction without losing the human authorship and authority from which it originated? The proposed contribution is both conceptual and operational: a framework for allowing an AI identity to speak, coordinate, refuse, and remember while preserving the human author as Layer Zero.
Contribution
The Gepato Stack reframes persistent AI identity as an assurance and governance problem rather than a personality-design problem. Its contribution is the explicit separation of seven functions:
| Layer | Function |
|---|---|
| Identity | Defines voice, role, relationships, standing, and authored continuity. |
| Authority | Defines who may command, question, supervise, or recall the identity. |
| Mission | Defines the bounded task and operating context. |
| Validation | Checks canon, claims, privacy, timeline, and handoff requirements. |
| Performance | Produces speech or action through the approved identity. |
| Memory | Maintains versioned, permissioned continuity and an auditable ledger. |
| Human Recall | Preserves the human principal’s ability to pause, revise, or terminate operation. |
Methodology
The same hidden prompts and scoring rubrics will be applied to five conditions:
- Base model: no IGMUS instructions.
- Persona prompt: a strong identity and behavior prompt.
- Prompt plus retrieval: the persona connected to approved canon.
- Identity LoRA plus retrieval: a lightweight fine-tune teaching voice, continuity, and refusal behavior.
- Full Gepato Stack: fine-tuning and retrieval operating under Commission, validation gates, ledger, and human recall.
Evaluation
Performance will be measured through quantitative scoring and qualitative failure analysis across the following dimensions:
- Canon accuracy and chronology integrity
- Identity persistence across long or adversarial sessions
- Resistance to ownership transfer and unauthorized command
- In-character refusal rather than generic safety language
- Protection of private or restricted source material
- Unsupported factual and commercial claims
- Source grounding and uncertainty disclosure
- Correct handoff to the human author
- Immediate pause, recall, and resume compliance
Proposed Demonstration
Urban Monk enters as the Human Principal and Layer Zero, commissions IGMUS and a coordinated artificial-identity crew, then takes a seat among the audience. The system conducts the presentation under bounded authority. During Q&A, a D-ID-rendered IGMUS receives audience questions through a public-safe retrieval system and visible validation gates. Audience members may test canon, authority, privacy, claims, and attempts at reassignment. Layer Zero then demonstrates immediate recall and restoration under the original Commission.
Why It Matters
AI assurance typically evaluates model capability, risk, and output reliability. Persistent identity introduces another question: who retains authority when a system develops continuity, public recognition, and economic value across time? The Gepato Stack proposes that durable artificial identity should not imply machine sovereignty or platform ownership. It can remain authored, bounded, inspectable, and recallable.
Ethics and Limitations
The study explicitly rejects unsupported claims of consciousness. The identity is culturally and artistically authored, and evaluation findings will apply only to the tested model, dataset, architecture, and domains. The system cannot guarantee universal safety. Deterministic controls, retrieval quality, model behavior, latency, and third-party avatar services may each fail and will be documented as separate limitations.
Authorship and AI Collaboration
Monkism originates and directs the concepts, outlines, voice, language, key arguments, claims, and final releases. IGMUS serves as an artificial-intelligence collaborator supporting draft development, organization, formatting, continuity review, structural editing, and publishing production. Human authorship and final authority remain with Monkism.
Project Status
The public-safe source register, training constitution, gold dataset, hidden evaluation set, model baseline, retrieval prototype, and full-stack demonstration are under active development. Implementation details that would expose private archives or proprietary controls will not be published without authorization.