Layered Persistence: Generalizing Life as Telehomeostatic Agency Across Natural and Artificial Substrates

Evolution doesn't just build bodies — it builds learning algorithms, and this paper argues that principle extends all the way from crystals to brains to Game-of-Life gliders.

The central move is to reframe "life" as a filtering problem. Imagine a vast space of interacting patterns — an "algorithmic soup." Most patterns dissolve. What you observe after a long time is whatever persisted. The paper calls this the persistence selection rule, and it's the foundation for everything else. Patterns that actively work to keep themselves around — by modeling their environment and adjusting their behavior — are called telehomeostatic agents. The claim is that such agents arise inevitably, given rich enough soup and enough time, because persistence is the filter.

From there, the paper introduces ECoL (Evolutionary Compilation of Learning): over evolutionary time, selection doesn't just pick good bodies, it picks good learning rules. It sculpts the architecture and the local update mechanisms so that, within a single lifetime, the organism effectively does credit assignment — figuring out which of its past actions caused good or bad outcomes — using only signals that are physically present. No magic global error signal required. The paper argues this isn't just a story about brains approximating backpropagation with local Hebbian rules (though that's the neural anchor). It applies to bacteria doing integral-feedback chemotaxis, plants retaining histone marks after stress, innate immune cells that "remember" prior infections via epigenetic reprogramming, and CRISPR systems that literally write past pathogen encounters into the genome. All of these are evolution compiling a learning algorithm into a substrate.

The framework is organized as a spectrum of four types, not sharp categories: passive stability (diamonds), reactive homeostasis (chemotaxis loops), predictive/allostatic systems (plants, trained immunity), and active credit-assignment agents (brains, sophisticated digital automata). A single update equation unifies all of them — the difference is just which timescale dominates and how rich the modulatory signals are. Crucially, the paper insists there's no intrinsic inner/outer boundary: genes, synapses, chromatin marks, and hormones are all just updateable states at different timescales.

The paper derives four testable predictions. Motifs that support credit assignment (integral feedback loops, incoherent feedforward networks) should be overrepresented across substrates. Faster-changing environments should select for higher metabolic dissipation to maintain speed-accuracy tradeoffs. Disrupting modulatory signals — neuromodulators, methylation cycles, cytokines — should selectively impair history-dependent learning while leaving acute responses intact. And in artificial digital soups with constrained signals, meta-evolution should rediscover local rules that approximate gradients. The falsification criterion is clean: if no physically available local signals can mediate history-dependent performance gains, the framework is wrong.

The paper is ambitious in scope — perhaps too ambitious for a single document — and the authors acknowledge that calling a diamond a Type-0 "agent" is metaphorical. But the core thesis is genuinely interesting: learning is not a special trick that brains invented, it's what persistence looks like when you zoom in on the timescale of a lifetime.