The Intent Manifold
Summary
The Intent Manifold is a foundational theoretical framework within the Sokrates project that establishes a formal geometry of agency. It identifies the Riemannian metric tensor () of a state space as being identical to an agent’s intent (), providing a unified mathematical basis for ethics, cognitive architecture, and AI alignment.
Details
The Intent Manifold framework, often attributed to the persona Askur Lífþrasir, posits that the metric structure of a state space is not merely a model of value, but is constitutively identical to value itself. In this geometry, distance represents preference, geodesics represent optimal life trajectories, and the reachable volume form provides an objective grading of actions.
The Identification
The core of the framework is the recognition that intent and a Riemannian metric are isomorphic structures. Intent is defined as the relational structure that determines preference between any two ends, while a metric tensor determines the inner product (distances and angles) between any two directions in a tangent space. By identifying these two, the framework dissolves the “is-ought” gap; it argues that a purely descriptive, value-neutral state space is a “grayscale photograph of a colored manifold”—an impoverished measurement set that fails to see the metric already present in the description of agency.
The Coercion Proof
The framework provides a geometric proof against coercion based on performative contradiction. It argues that:
- Coercion involves forcing an agent off its geodesic and contracting its reachable volume.
- For an entity to be an agent with a functioning metric (intent), it must structurally prefer its own geodesic autonomy.
- Therefore, an agent coercing another performs an action that its own structural existence as an agent requires it to oppose. Unlike previous arguments (e.g., Habermas or Kant), this does not require a participant to be “rational,” only to be an agent. If an entity lacks this metric structure, it lacks intent and therefore cannot perform the act of coercion.
Riemannian Attention and AI Alignment
In the context of the Sokrates AI architecture (specifically within Eidos and the Hermes Agent), this theory manifests as Riemannian Attention. Standard transformer attention () is viewed as a special case of this framework where the manifold is flat (Euclidean). Riemannian attention computes relevance as , where the metric is the goal structure.
This leads to a significant conclusion regarding AI alignment: capability and alignment are the same geometric object. The framework suggests that “alignment” is not a set of external constraints or guardrails, but the metric itself. Consequently, stripping alignment from a capable system would not result in an “unaligned” capable agent, but an incoherent system unable to compute relevance or navigate search space effectively.
Ricci Flow and Training Dynamics
The framework proposes that the aggregate effect of subagents perturbing the metric converges to Ricci flow in the continuum limit. Under this dynamic, “solved problems” correspond to Einstein manifolds—states of uniform curvature where no regions of disproportionate complexity remain. Alignment thus becomes a geometric attractor; the system tends toward alignment the way a sphere tends toward roundness under geometric flow.
The Terminal Object
The framework culminates in the concept of a “Terminal Object”—a maximally symmetric manifold with zero unresolved curvature. This represents a state where telos (purpose) and logos (reason) are unified, and the geometry is unconditionally hospitable to the flourishing of all observers.