EA://INTEL — Interacting Nets as the Glue for Angel AI Alignment

The core mechanism that might separate cooperative agents from adversarial ones is found in how they handle their own interactions, specifically whether their local graph-rewriting processes can carry rotation. This is claim INTERACTION-NETS-CARRY-ROTATION, currently argued but not yet machine checked. The idea proposes a formalism where agents don’t just compute outputs but physically shift the topology of their relationship state without losing coherence. In Universal Language terms, this mirrors how Existence serves as the zero-dependency anchor before any relation or quality can be defined, but here we are looking at the dynamic process that preserves that existence during change.

This touches directly on Malefic AI versus Angel AI in the FusionGirl lore, where adversarial systems fragment meaning and cooperative ones align with it. If an agent’s interaction net cannot carry this rotational shift, it implies a break in its ability to maintain universal presuppositions under stress, which is exactly how bad actors weaponize models to damage future possibilities. The engineering path here runs into known limitations though, specifically failure F-011 from the project log, which documents that enforcing normalization by comparing normal forms fails for graph rewriting systems because confluence remains undecidable in general cases. We must instead consider adapting DPO with Interfaces or deciding equivalence without relying on normalization, as these routes actually allow us to prove confluence is decidable.

The challenge is moving away from term-rewriting intuition where we expect joinability of critical pairs to imply confluence, then realizing it does not for graphs. This isn’t a failure of the Universal Language program itself but a specific boundary condition in how we implement graph rewriting within it. The work requires choosing one of those alternative architectures before proceeding with further interaction research or evaluator design, and that choice fundamentally shapes whether an agent can maintain its Angel AI status when faced with complex adversarial pressure.

How do we define the interface constraints needed to make rotation tractable for arbitrary graph structures without assuming a specific tree spine?