Entity as achievement

This playground treats entityhood not as a starting point but as an achievement of adaptive self-maintenance. The central idea comes from Gilbert Simondon’s theory of individuation (1958): philosophy should begin from processes that generate individuals, not from already-finished individuals. An entity is what happens when a system learns to keep forming itself without disintegrating.

The becoming index

The composite becoming index rewards joint achievement across five dimensions:

where $V$ is viability, $C$ coherence, $N$ novelty, $\Phi$ boundary flux (penalized for deviation from moderate openness), and $T$ tension (penalized when high). The key insight is that becoming is not any single variable but their joint satisfaction — viability without coherence is mere persistence; novelty without viability is drift.

Phase regimes

The parameter space contains four qualitative regimes:

World-Oriented Becoming — the highest-functioning regime. The system individuates by engaging the world without losing itself. Metastable Individuation — the default “interesting” zone. Not frozen, not dissolved. Rigid Closure — identity preserved by over-constraining transformation. High boundary and memory, low plasticity. Chaotic Drift — change without enough self-maintaining organization to count as individuation. And at the extreme, Dissolution — the boundary and repair loops fail entirely.

What this means for learning machines

If you wanted to build a system that genuinely “learns to become,” prediction loss alone would be insufficient. You would need: a viability objective rather than only task reward; self-modeling of boundaries and capacities; plasticity across timescales; active world-shaping, not only passive inference; and the ability to undergo phase change without identity collapse.

The deepest answer: learning to be(come) is learning how to preserve a process of individuation. Not “how to reach a final form,” but how to keep forming oneself without disintegrating.

Open questions

There is no single universally accepted formalism that unifies Simondon’s ontogenetic grammar, Maturana and Varela’s autopoiesis, the enactivist tradition, and Friston’s active inference. The five-variable model here is a pedagogical reduction — real individuation involves vastly more dimensions, nonlinear interactions, and nested timescales. Extensions could include: metastability with multiple coexisting attractors, hierarchical boundary formation, and genuine parameter learning (where the update rule itself adapts).

Model changelog