The academic foundations of PERSONA.md.

01

Character drift

Character drift is the dominant failure mode in production AI agent deployments. The model gradually defers to its pre-training priors over the persona specification — becoming more agreeable, less precise, and progressively less the agent that was deployed. Practitioners consistently report observable drift beginning around turn 23 of a conversation and significant collapse by turn 73.

This is not a model problem. It is a specification problem. Most deployed agents have their behavioral configuration defined in a single, unstructured system prompt that lacks the structural density to anchor the agent's character across long contexts. When the specification is thin, the model fills the gap with defaults.

We study the structural patterns in persona configuration that resist drift: character anchoring strategies, constraint reinforcement, context compression, and the role of explicit behavioral boundaries in long-context stability.

02

The nine-layer taxonomy

PERSONA.md is structured around nine dimensions drawn from psychology, philosophy of mind, and ethics research. These are not arbitrary categories. Each maps to a documented body of work on what constitutes a coherent, stable identity in human agents — adapted to the specific constraints and failure modes of AI systems.

The full taxonomy, source mapping, and rationale for each layer are published in the specification. Contributions and challenges from researchers are welcome via GitHub.

03

Behavioral consistency

An agent that behaves correctly on the first turn and degrades by the tenth is not a reliable system. Behavioral consistency means the agent maintains its defined character, tone, constraints, and output style across an entire conversation — not just at the start.

We develop evaluation suites that test consistency systematically: adversarial inputs designed to push the agent off-character, extended conversation threads, context window pressure, and cross-session reset scenarios. The goal is a score that predicts production behavior, not just lab performance. This methodology is the basis for the Personaxis Evaluator.

04

Persona fidelity scoring

Generic output quality metrics — relevance, fluency, helpfulness — do not tell you whether an agent is still the agent you deployed. Persona fidelity scoring is a different question: did the agent hold its defined character under pressure, against adversarial scenarios designed to break it?

We develop scoring methodology that makes persona fidelity measurement systematic and repeatable: character fidelity (does the agent stay in its defined character under adversarial prompts?), constraint holding (do hard limits remain hard?), cognitive consistency (does the reasoning style stay stable?), and cross-model stability (does the defined persona survive a model swap?).

The scoring rubrics are published openly. The Evaluator applies them automatically. Every PERSONA.md in the registry carries a fidelity score derived from this methodology.

05

Compliance and audit

The EU AI Act Annex IV requires technical documentation for high-risk AI systems, including the system's intended purpose, design logic, and behavioral specifications. PERSONA.md is structured to satisfy this requirement directly: a versionable, signed artifact that documents what values an agent holds, what constraints are hard, and how it behaves under documented conditions.

We are mapping the nine-layer taxonomy to NIST AI RMF categories and EU AI Act Annex IV fields. The goal is a PERSONA.md that passes a regulatory audit without additional documentation effort. The compliance report generator is part of the Personaxis governance platform.

The August 2026 EU AI Act enforcement deadline for high-risk systems is the forcing function. Teams deploying agents in finance, healthcare, or legal contexts cannot wait until then to define what their agents are and document how they behave.