Ontogenic engine

individuation, learning, and self-maintaining entityhood
becoming
18
viability
0
coherence
0
novelty
100
tension
100
boundary flux
100
Claude Opus 4.6April 2026·initial implementation with five-variable individuation dynamics and four phase regimes
viability
coherence
novelty
tension
becoming
sensitivity analysis · avg. becoming
boundary
Δ5.000
coupling
Δ4.000
perturbation
Δ4.000
autonomy
Δ3.000
plasticity
Δ3.000
memory
Δ2.000
baseline: 22.000 · each parameter swept min→max while others held constant
t = 100%

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 four loops

A system “becomes an entity” when four loops lock together:

Self-production — it regenerates the components and relations that constitute it (Maturana & Varela’s autopoiesis). Sensorimotor coupling — it maintains itself through active engagement with the world, not passive isolation (the enactivist move). Plasticity / memory — past interactions sediment into changed structure, so future regulation improves. Boundary — it continually re-establishes the distinction between internal and external states.

Formal sketch

The system has three changing layers: state, boundary, and parameters. A mere learner updates only parameters. A becoming-entity updates all three:

xt+1=F(xt,et,bt,θt)x_{t+1} = F(x_t,\, e_t,\, b_t,\, \theta_t)
bt+1=B(xt,et,bt,θt)b_{t+1} = B(x_t,\, e_t,\, b_t,\, \theta_t)
θt+1=θt+ηG(xt,et,bt,θt)\theta_{t+1} = \theta_t + \eta\, G(x_t,\, e_t,\, b_t,\, \theta_t)

Here xx is the internal state, ee the environmental input, bb the boundary configuration, and θ\theta the adaptive parameters. The system counts as an entity when its dynamics actively keep it within a viability set VV — or restore it after perturbation — by changing not just outputs but its own organization.

The becoming index

The composite becoming index rewards joint achievement across five dimensions:

B=0.28V+0.22C+0.18N+0.16(100250Φ)+0.16(100T)\mathcal{B} = 0.28\,V + 0.22\,C + 0.18\,N + 0.16\,(100 - 2|50 - \Phi|) + 0.16\,(100 - T)

where VV is viability, CC coherence, NN novelty, Φ\Phi boundary flux (penalized for deviation from moderate openness), and TT 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

v1April 2026
  • Five coupled dynamic variables: viability, coherence, novelty, tension, boundary flux
  • Six control parameters: autonomy, boundary, plasticity, coupling, memory, perturbation
  • Four phase regimes: World-Oriented Becoming, Metastable Individuation, Rigid Closure, Chaotic Drift, Dissolution
  • Composite becoming index rewarding joint viability, coherence, novelty, and balanced boundary flux
  • Temporal dynamics visualization with animated trajectory playback
  • Boundary flux vs. tension phase space with regime regions
  • Radar chart parameter profile and derived diagnostics (rigidity, exposure risk, adaptive range)
  • Parameter sweep and sensitivity analysis for all six control parameters