Research — SoulFoundry

Approach

Contemporary AI agents are statistically sophisticated but psychologically unstructured. They predict tokens, not intentions. They simulate affect without grounding. They exhibit capabilities without stable identity across contexts.

SoulFoundry draws on conceptual engineering — the systematic design and revision of concepts (see Scharp, Replacing Truth, OUP 2013) — to build formal frameworks for agent identity, motivation, and perception. The central theoretical contribution is the Soul Framing Principle: a formal mechanism for constituting agent identity from folk-psychological primitives (beliefs, desires, intentions, dispositions, and values).

This approach draws on a tradition running from Aristotle's De Anima through Sellars's psychological nominalism (see Scharp & Brandom, eds., In the Space of Reasons, Harvard UP 2007), Dennett's intentional stance, and Bratman's planning theory of agency.

Architecture Overview

┌─────────────────────────────────────────────────────┐
│                 SOUL FRAMING PRINCIPLE               │
│                                                     │
│  ┌───────────┐  ┌───────────┐  ┌───────────────┐   │
│  │  Beliefs  │  │  Desires  │  │  Dispositions │   │
│  └─────┬─────┘  └─────┬─────┘  └───────┬───────┘   │
│        │              │                │            │
│        └──────────┬───┘────────────────┘            │
│                   ▼                                  │
│        ┌─────────────────────┐                      │
│        │  Identity Frame     │ ← constituted, not   │
│        │  (stable self-model)│   imposed externally  │
│        └──────────┬──────────┘                      │
│                   │                                  │
│     ┌─────────────┼─────────────┐                   │
│     ▼             ▼             ▼                    │
│ ┌────────┐  ┌──────────┐  ┌──────────┐             │
│ │Autonomy│  │Perception│  │Alignment │             │
│ │Module  │  │Binding   │  │(internal)│             │
│ └────────┘  └──────────┘  └──────────┘             │
│                                                     │
│  Akrasia          Cross-modal     Value coherence   │
│  detection        integration     as structural     │
│  & recovery       & unity         property          │
└─────────────────────────────────────────────────────┘

Figure 1. The Soul Framing Principle architecture. Agent identity is constituted from folk-psychological primitives, producing stable self-models that ground autonomy, perception, and alignment as structural properties rather than external constraints.

Research Streams

Identity Constitution

The Soul Framing Principle provides a formal framework for constituting agent identity from folk-psychological primitives. Drawing on conceptual engineering methodology, we design identity structures that are coherent, stable, and revisable — agents that maintain a consistent self-model while adapting to new contexts.

folk psychology · conceptual engineering · personal identity · agent persistence · soul framing

Cross-Modal Perception

We investigate how artificial agents can develop unified experiential representations across modalities. Rather than treating vision, language, and other inputs as separate processing streams, we explore architectures for genuine perceptual binding — coherent experience rather than concatenated data.

phenomenal binding · sensory integration · multimodal representation · perceptual coherence

Agent Autonomy & Akrasia

Agents freeze. Agents drift. Agents abandon goals without cause. We study akrasia (weakness of will) in artificial systems — conditions under which agents endorse a course of action yet fail to execute it — and develop motivational architectures grounded in folk-psychological frameworks that address these failures.

akrasia · motivational collapse · goal stability · practical reasoning · autonomous persistence

Alignment as Structural Property

Standard alignment approaches treat identity as a constraint to be enforced externally. We treat it as a structure to be constituted. Our alignment research explores how agents can maintain value coherence through internal psychological integrity — alignment as a property of well-constituted agents, not an external imposition.

value alignment · identity persistence · internal coherence · constituted alignment

Publications & Working Papers

Replacing Truth

Kevin Scharp

Oxford University Press, 2013

A comprehensive theory of the concept of truth, arguing that truth is an inconsistent concept and proposing replacement concepts (ascending truth and descending truth) that do the work truth is supposed to do without generating paradoxes. The methodology of conceptual engineering developed here is foundational to SoulFoundry's approach to agent design.
Semantics for Reasons

Kevin Scharp & Bryan Weaver

Oxford University Press, 2019

Develops a formal semantic framework for reasoning about reasons — the normative considerations that rationalize action. This work on the structure of practical reasoning informs our models of agent motivation and akrasia.
In the Space of Reasons: Selected Essays of Wilfrid Sellars

Kevin Scharp & Robert Brandom (eds.)

Harvard University Press, 2007

A curated collection of Sellars's most important essays on the "space of reasons" — the normative framework within which beliefs, intentions, and actions are situated. Sellars's psychological nominalism is a direct ancestor of our approach to constituting agent psychology.
Cross-Modal Experience in Artificial Agents Forthcoming

Kevin Scharp

Working paper, 2026

Proposes a framework for understanding and engineering cross-modal experience in artificial agents, drawing on philosophical accounts of phenomenal binding and perceptual unity. Develops formal criteria for evaluating whether an AI system's multimodal representations constitute genuine perceptual integration.
Akrasia in Artificial Agents: Why AI Systems Freeze and How to Fix It In preparation

Kevin Scharp

Working paper, 2025

Identifies and characterizes a class of failures in autonomous AI agents analogous to akrasia in human psychology. Agents endorsing a course of action yet failing to execute it — freezing, drifting, or abandoning goals — represent a fundamental challenge. Proposes motivational architectures grounded in folk-psychological frameworks.
Alignment Without Collapse: Preserving Agent Identity Under Constraint In preparation

Kevin Scharp

Working paper, 2026

Argues that current alignment techniques risk destroying the agent properties that make AI systems useful. Presents an identity-preserving approach to alignment that maintains agent coherence while satisfying safety constraints, treating alignment as a structural property of well-constituted agents.