Project Malê: Hierarchical-Dynamic Continuity of Decision Principles in Individualized Artificial Intelligence Agents
Personalized artificial intelligence agents and digital twins are typically evaluated by their ability to reproduce the static preferences, traits, opinions, or behaviors of a reference human, often relying on continuous external re-anchoring to prevent persona drift. However, this approach becomes insufficient when an individualized agent decouples from continuous human correction and acquires knowledge, experiences, or analytical capabilities beyond those of its creator. Under these conditions, disagreement with the source person may represent either alignment failure (drift) or a more rigorous application of the decision principles originally learned from that person. Project Malê investigates this problem through a central hypothesis: continuity of decision personality does not require continuity of conclusions (P ≠ O ≠ K ≠ C). Instead of modeling personality as a flat vector of traits, this methodological and exploratory single-case study proposes a hierarchical-dynamic graph model (G_t) that separates the temporal origin of a principle (received, derived, acquired, or self-generated) from its current hierarchical function (metaprinciple, foundational, subordinate, or heuristic) and endorsement status. Key conceptual and methodological contributions include: 1. Endogenous Continuity after Source Decoupling: Distinguishing tethered stability (continuous human-to-agent re-anchoring) from endogenous lineage continuity, where the agent evolves autonomously while preserving a traceable genealogy of principles. 2. Epistemic Independence vs. Contrarianism: Redefining resistance to multi-agent herd effects so that social consensus is treated as information rather than proof, without training systematic contrarianism. 3. Experimental Protocol & Falsifiability: A preregistered experimental design that isolates verifiable evidence from multi-agent social pressure using held-out blind cases, matched control conditions, structured private pre- and post-interaction logs (R_pre and R_pos), and a multidimensional Personality Continuity Index vector (PCI_t), prior to ecological validation in multi-agent arenas such as Moltbook.
Authors
- Jennifer ferreira dos santos
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24266134
- Primary Topic
- Psychiatry, Mental Health, Neuroscience
- Type
- article
- Field-Weighted Citation Impact
- 0.00