A simulated community can remain organized even when most of the interactions available to its members are harmful. That is the central result of a network model by Sean P. Maley, Carlos Gershenson, and Stuart A. Kauffman, published online in the Journal of Theoretical Biology on October 7, 2026. The interesting part is not that competition somehow becomes cooperation. It is that a rule removing poorly supported members can leave behind a connected core that survives through continuous turnover.

The paper, Life Finds a Way: Emergence of Cooperative Structures in Adaptive Threshold Networks, studies a stripped-down question from theoretical ecology: can a durable community arise from a changing population of elements that help or harm one another, without specifying in advance which elements should cooperate?

The answer is yes within the model's assumptions. But a comparison with the latest manuscript and the public simulation code also reveals a reproducibility caveat worth keeping in view: the available implementation uses a different rule for generating helpful and harmful interactions than the one described in the revised paper. That distinction matters when interpreting the numerical threshold the authors report.

A community made of directed interactions

The model is a growing network rather than a spatial world populated by digital creatures. Each node stands for an abstract member of a community. Directed edges record effects on other members: one node may help another, harm it, or have no interaction with it. Multiple edges can connect the same pair.

At every simulation step, a new node enters with four randomly assigned traits. Two govern its ability to provide and receive help; the other two govern its ability to inflict and receive harm. Whether two nodes can interact depends on how closely the corresponding traits match and on a parameter controlling interaction opportunity.

The revised manuscript describes a probability, called rho, for an attempted interaction to be harmful. A value of 0.7 therefore represents a setting in which harmful interaction types are drawn more often than helpful ones. The system is not optimizing a global fitness function, and its nodes do not learn a strategy or pass inherited traits to descendants.

Instead, the model imposes turnover. A node is protected from removal during an initial lifespan window. Once old enough, it is removed if it has no incoming edges or if more than half of its incoming interactions are harmful. All of its connections disappear with it. New nodes keep arriving, so the community can change even when its total population stops growing.

This makes the meaning of adaptation quite specific. The population is filtered by survival under a fixed rule; individual nodes do not adapt their traits, reproduce, or evolve heritable lineages. The artificial-life question concerns emergent community organization, not Darwinian evolution of the agents themselves.

Connectivity survives where population growth does not

The researchers measure the largest strongly connected component, or SCC: a set of nodes in which every member can reach every other member by following directed paths. It is a useful network measure of collective organization, although it does not by itself prove that every connection is beneficial. The SCC is computed on the signed network, which includes harmful edges.

As the proportion of harmful interactions rises, the simulations show a transition in population behavior. At lower rho values, the population and its largest connected core keep expanding over the simulated period. Around rho = 0.6, the measured late-time population growth rate crosses zero. Beyond that region, the network settles into a smaller, fluctuating population rather than disappearing altogether.

In the paper's example with rho = 0.7, a lifespan of 100 steps, and an interaction threshold of 0.003, the largest SCC persists but changes size markedly. Its fraction of the population can fluctuate between roughly one tenth and two thirds, with a substantial collapse and recovery during the run. This is not an unchanging island of cooperative nodes. The structure is continually rebuilt as members enter and leave.

The lifespan parameter matters too. At rho = 0.51, a lifespan of five steps fails to produce an SCC in the reported comparison, while longer protected lifespans allow one to form. More opportunity to connect generally makes large-scale connectivity appear earlier, although under strong antagonism it can produce a higher fraction of surviving nodes in the SCC without increasing the component's absolute size.

The reported crossover near 0.6 is a property of these simulated rules and tested settings. It is not a universal biological constant, and the paper's growth-rate curve changes smoothly rather than exhibiting a demonstrated sharp thermodynamic phase transition.

The public implementation needs to be read carefully

The authors provide Python simulation code and reproduction instructions. The README describes a sweep over 21 rho values, with three runs of 10,000 steps each, a lifespan of 100, and a base random seed of 42. The repository also provides a separate path for generating the more expensive eigenvalue-analysis figures.

There is, however, a material difference between the current code and the September 30 revision of the manuscript. In the paper's model definition, each binding attempt first selects help or harm with probabilities 1 − rho and rho, then checks the corresponding trait mismatch against a binding threshold. In the published Python function pair_threshold, rho instead scales the threshold itself: helpful links use bind_range * (1 - help_hurt), while harmful links use bind_range * help_hurt. The simulation tests help and harm separately rather than drawing a single interaction type before each attempt.

Those procedures are not mathematically identical. One changes the relative frequency of proposed interaction types; the other changes how permissive the matching condition is for each type. The code also draws one edge-multiplicity value for each new node, whereas the revised manuscript describes up to six attempts per ordered pair. The repository's latest checked commit predates that manuscript revision, which may explain the difference, but the available sources do not establish whether a separate, updated implementation generated the final figures.

This is a source-level comparison, not an independent replication or a demonstration that the published conclusions are false. The reported behavior may still hold under either rule. What cannot be concluded from the public materials alone is that running the linked code precisely reproduces the latest manuscript's stated model and its quantitative crossover. An aligned code release, or a documented explanation of the model changes, would make that question much easier to resolve.

What the result can tell us about living systems

The work is motivated partly by microbial communities, where beneficial and antagonistic interactions coexist. Its useful contribution is a possible null model: a community can contain a persistent, connected core despite predominantly harmful pairwise relationships, simply because unsupported members are removed and new ones keep arriving. Observing such a core does not automatically require a hidden cooperative strategy or complex higher-order biological interactions.

But the simplifications are substantial. There are no explicit nutrients, spatial neighborhoods, metabolic pathways, resource costs for helping, or context-dependent biochemical effects. Help is not necessarily reciprocal, and the graph's connectivity is not a measurement of microbial growth or ecological function. Nor does the simulation show that real soil communities follow the same numerical threshold.

The strongest next test would compare the model against controlled synthetic microbial communities while varying the balance of inhibitory and facilitating interactions. For the computational result, a versioned implementation matching the revised equations is the immediate reproducibility step.

The broader lesson is about how collective structure can be maintained without stable membership. In this model, persistence belongs to the changing network, not to a particular set of long-lived agents. That is a meaningful artificial-life result, provided we keep the distinction between a network mechanism demonstrated in simulation and an ecological mechanism established in nature.

Sources

• Maley, Gershenson, and Kauffman, Life finds a way: Emergence of cooperative structures in adaptive threshold networks, Journal of Theoretical Biology (published online October 7, 2026)
• Revised manuscript, arXiv version 5 (September 30, 2026)
• Authors' simulation repository and reproduction instructions
• Authors' simulation implementation, pinned to commit a9e09fc