Phenopoiesis Algorithm Overview

• Phenopoiesis Algorithm is a computational framework that implements phenotype-first evolution by integrating genetic and epigenetic inheritance systems.
• It enables rapid adaptation, minimized catastrophic forgetting, and improved multi-task performance, outperforming classic gene-centric and Baldwinian models.
• The algorithm uses a two-timescale process—ontogenetic trials for immediate learning and generational mutations for long-term evolution—to reinforce effective phenotypic patterns.

The Phenopoiesis Algorithm defines an algorithmic framework for phenotype-first evolution, operationalizing organismal agency as the heritable transmission of learned phenotypic patterns across generations. Unlike classical gene-centric evolutionary models—which attribute causality exclusively to genetic inheritance—the Phenopoiesis Algorithm introduces a dual inheritance system comprising both genomic (gene-based) and epigenomic (phenotypic-pattern) channels. This approach quantitatively demonstrates the adaptive value of inheriting phenotypic compositions discovered during lifetime learning, resulting in significantly accelerated adaptation, reduced catastrophic forgetting, and enhanced multi-task capability compared to standard gene-centric or Baldwinian (within-lifetime learning only) populations (Le, 1 Feb 2026).

1. Formalization and Core Constructs

The algorithm operates in a domain of $10 \times 10$ binary grids, with a finite target set of shapes $\mathcal{T}$ (e.g., L, T, Plus, Cross, Square). Each organism encodes its phenotype as a grid $p \in \{0,1\}^{10 \times 10}$, constructed by composing primitives from a fixed library $\mathcal{C}$. A composition recipe $r$ is an ordered tuple of primitives with placement instructions; composing $r$ yields a phenotype candidate.

Organisms inherit both: (i) a genome $G$, a bit-string or real-valued vector indicating primitive sampling propensities, and (ii) an epigenome $E$, a set of composition recipes discovered during ontogenetic (within-lifetime) trials. Genetic inheritance proceeds via mutation: $G^{(g+1)} = \mathrm{Mutate}(G^{(g)})$. Phenotypic inheritance transmits $E$ directly: $E^{(g+1)} = E^{(g)}$. Fitness is computed as the maximum translation-invariant Intersection-over-Union between phenotype and target, $f(p, t)$.

This dual-channel inheritance demarcates the genome as a slowly evolving, mutable substrate for primitive selection, while the epigenome functions as a high-capacity, rapidly-adaptive memory of successful solutions (Le, 1 Feb 2026).

2. Theoretical Structure: Two-Timescale Dynamics

The Phenopoiesis Algorithm implements two distinct timescales:

• Fast (Ontogenetic) Timescale: During each lifetime, an organism conducts $T$ developmental trials. In each, it probabilistically decides to exploit (reuse $r \in E$) or explore (sample new $r$ via $G$ and $\mathcal{C}$). Upon discovering higher-fitness phenotypic recipes, it writes these into $E$ and reinforces $G$ on the successful primitives/placements.
• Slow (Evolutionary) Timescale: At the generational boundary, parents are selected for reproduction in proportion to their best trial fitness; genomes of offspring are mutated, and epigenomes are inherited unaltered.

Explicit inheritance equations are: $$G_\mathrm{offspring} = G_\mathrm{parent} \oplus \mathrm{mutation}, \qquad E_\mathrm{offspring} = E_\mathrm{parent}$$ Thus phenotypic memory persists undisturbed, while genetic channels continue to introduce slow exploration.

This two-scale division directly implements Noble's phenotype-first causality by allowing discovered phenotypic structures to write back to offspring, in contrast to gene-centric or Baldwin control procedures that limit transmission to genomic channels only (Le, 1 Feb 2026).

3. Algorithmic Steps and Pseudocode

A concise pseudocode representation follows:

Input: Population N, max generations G_max, trials T, primitives C Initialize population: for each i=1..N, organism O_i^0 = (G_i^0, E_i^0 = ∅) For generation g = 0 .. G_max-1: For each organism O_i^g = (G, E): f_best ← 0; r_best ← ∅ For t = 1..T: With probability α: r ← sample from E (exploit) Else: r ← sample new from G and C (explore) p ← Compose(r) f ← fitness(p, target) If f > f_best: f_best ← f; r_best ← r E ← E ∪ {r}; G ← reinforce(G, r) Save f_best for selection Update O_i^g with modified (G, E) Select N parents by tournament on f_best For each offspring: G ← Mutate(G_parent) E ← E_parent (no mutation)

Mechanistically, lines implementing write-back (epigenome updates) distinguish the algorithm from all gene-only or Baldwinian procedures (Le, 1 Feb 2026).

4. Computational and Memory Analysis

The dominant computational cost per generation is $O(N \, T \, S^2)$, where $N$ is population size, $T$ lifetime trials, and $S$ grid side (10). For comparison, gene-only (GENE) models incur $O(N S^2)$. Epigenome storage grows as $O(|E| \cdot k)$, with $k$ primitives per recipe and $|E|$ at worst $G_\mathrm{max} \cdot T$, though practical implementations cap $|E|$ (e.g., retain top $M$ patterns).

Epigenome exploitation (sampling from $E$) is $O(1)$ if array-based but can approach $O(|E|)$ in the worst case. Memory requirements are dominated in practice by bounded epigenome buffers, typically $O(M \cdot k)$ per organism (Le, 1 Feb 2026).

5. Experimental Evaluation and Quantitative Results

Experimental setups utilize $N=50$, $T=20$, $G_\mathrm{max}$ variable (100–500), and five target shapes on $10\times10$ grids. Statistical analysis includes ANOVA with Tukey post-hoc, $p<0.05$, over 30 random seeds.

Key adaptation metrics:

• Generations to 80% fitness (target switch):
  • GENE: $28.4\pm8.9$; BALDWIN: $14.7\pm5.2$; PHENO: $8.3\pm3.1$
  • $3.4\times$ faster adaptation with PHENO versus GENE; $1.8\times$ versus BALDWIN
• Catastrophic forgetting (rapid switching, 20 gen/task):
  • GENE: 51.1%, BALDWIN: 36.3%, PHENO: 31.1%
• Multi-task simultaneous fitness (L, T, Plus):
  • GENE: $70.0\%\,(\sigma=14.5\%)$
  • BALDWIN: $80.1\%\,(\sigma=15.8\%)$
  • PHENO: $91.2\%\,(\sigma=6.8\%)$

These results empirically support the claim that phenotypic inheritance confers substantial sample efficiency, rapid adaptation, reduced forgetting, and robust multi-task composition, with statistical significance ($p<0.001$ for PHENO vs GENE; $p<0.005$ for PHENO vs BALDWIN) (Le, 1 Feb 2026).

6. Comparative Mechanistic Analysis

The Phenopoiesis Algorithm is contrasted to both gene-only (GENE) and Baldwinian (BALDWIN) controls:

Feature	GENE	BALDWIN	PHENO
Inheritance	Genetic only	Genetic only	Genetic + Epigenetic
Lifetime trials	1	20	20
Write-back	none	none	genome & epigenome
Causality	G → P	G → P	P→E, G→offspring (bidirectional)
Exploration	global mutation	within-lifetime	within-lifetime + pattern reuse
Multi-task cap.	low	moderate	high (episodic memory)
Sample efficiency	low	medium	high

• GENE encodes only static, one-way transmission (G → P), with no learning or memory.
• BALDWIN permits within-lifetime learning but discards all acquired information intergenerationally.
• PHENO’s dual inheritance and write-back to $E$ instantiate phenotype-first, bidirectional causality as postulated in organismal agency frameworks (Le, 1 Feb 2026).

Adaptive advantages uniquely realized by PHENO include rapid re-adaptation through inherited phenotypic templates, minimized forgetting via population-level episodic memory, and compositional multi-task capacity without need for multi-objective optimization.

7. Connection to Broader Contexts and Related Algorithms

The Phenopoiesis Algorithm provides a concrete algorithmic realization of Denis Noble’s “phenotype-first” theory, rendering previously philosophical constructs of organismal agency in computationally testable form. A plausible implication is that algorithms incorporating explicit, heritable records of learned phenotypic solutions may offer broadly applicable advantages in non-stationary and multi-task environments beyond the grid-world demonstration arena. In cellular systems, the general concept of phenopoiesis is also addressed in the context of lineage analysis by the Lineage EM (LEM) algorithm, which infers heritable latent cellular states from division-time decorated lineage trees using EM and survivorship bias correction (Nakashima et al., 2018), suggesting that algorithmic phenopoiesis has relevance across both developmental biology and computational evolutionary dynamics.