Surviving by Serving (SBS) in Systems

• Surviving by Serving (SBS) is a principle where components persist by actively contributing functionally, enabling self-organization without explicit global objectives.
• SBS is applied across domains—such as network survivability, operating system security, and self-training architectures—to maintain functionality through proactive provisioning and constrained recovery.
• The SBS framework leverages utilization-dependent credit mechanisms and dynamic adaptation to drive emergent organization and pre-adaptive exploration in complex, multi-agent environments.

Searching arXiv for "Surviving by Serving" and closely related usages to ground the article in current papers. Surviving by Serving (SBS) denotes a class of ideas in which persistence is coupled to functional contribution rather than to centralized control, dense reward, or purely exogenous fitness. In its most explicit formulation, SBS is a general principle of self-organization: components persist as long as their outputs are utilized by other components, whereas prolonged non-utilization promotes adaptation and exploration (Metzner et al., 25 Jun 2026). Closely related operationalizations appear in SDN/NFV survivability, where service chains remain operational through proactively provisioned shared backup VNFs (Aidi et al., 2018); in operating-system security, where compromised services are restored and then constrained to continue only their core functions in a degraded mode (Chevalier et al., 2019); and in self-training architectures, where behaviors are retained only if their environmental consequences persist and preserve future interaction, although that framework is formally named “Sustainable Self-Training Through Environment-Mediated Selection” and “negative-space learning (NSL)” rather than SBS (Dodgson et al., 18 Jan 2026).

1. Definition and terminological scope

In the canonical SBS formulation, the local rule is simple: utilized outputs induce persistence or stabilization, while non-utilized outputs induce adaptation or exploration. The significance of the proposal lies in its claim that organized structure can emerge without a global objective, centralized supervision, or explicit optimization; functional utilization itself acts as a distributed selection principle (Metzner et al., 25 Jun 2026).

Across adjacent literatures, the phrase is used less as a universal theory and more as an engineering logic. In NFV, the relevant intuition is that service chains survive by continuing to serve traffic through pre-provisioned backups rather than by waiting to recover after failure (Aidi et al., 2018). In intrusion tolerance, the corresponding idea is that a service survives by continuing to serve its core functions under selective restrictions, rather than by full shutdown or machine-wide quarantine (Chevalier et al., 2019). In self-training, the closest equivalent is environment-mediated selection: behaviors survive only when they produce persistent positive environmental consequences that preserve future operability (Dodgson et al., 18 Jan 2026).

A recurrent source of confusion is acronymic rather than conceptual. In astronomy, “SBS” commonly denotes the Second Byurakan Survey, a galaxy survey that followed the First Byurakan Survey and used UV-excess and emission lines as selection criteria; that usage is unrelated to Surviving by Serving (Hakopian, 2014).

Context	What persists	Operational criterion
Complex adaptive systems	Components or agents	Outputs are utilized by other components
Self-training	Behaviors or trajectories	Positive, persistent environmental impact
SDN/NFV	Service chains	Shared backups sustain chains after node failure
OS security	Services	Core functions remain available in degraded mode

2. Local mechanism in the canonical SBS model

The minimal SBS model is a multi-agent interaction network with a shared state space and four active components: a raw-material source, an aging mechanism, transformation agents, and evaluators. The simulation runs for $N_{stp}$ discrete episodes; at the beginning of each episode, the system injects two raw-material items, one of each of two fixed raw-material types, while existing states can disappear with fixed probability $P_{dis}$ (Metzner et al., 25 Jun 2026).

States are represented as feature vectors in a continuous $N_{dim}$-dimensional space with $N_{dim}=5$, normalized to unit length so that they lie on a hypersphere. Compatibility is measured by cosine similarity,

$S(\vec{X},\vec{Q})=\vec{X}\cdot\vec{Q},$

and a state is suitable for a requirement when $S(\vec{X},\vec{Q}) \ge \Theta_{thr}$. Transformation agents search for two compatible input states, consume them, and generate two copies of an output state. The production rule is

$$\vec{Y}_i = \operatorname{norm} \left[ \operatorname{norm} \left( a_i \vec{X}^{(A)}_i + b_i \vec{X}^{(B)}_i \right) + \eta_{oos} \operatorname{norm}(\vec{R}_i) \right].$$

The out-of-subspace contribution $\eta_{oos}\operatorname{norm}(\vec{R}_i)$ is crucial because a pure linear combination of the two inputs remains in their span; the extra term permits novelty generation beyond the initial subspace (Metzner et al., 25 Jun 2026).

The SBS rule is implemented through utilization-dependent credit and adaptation. An agent receives one credit point when one of its outputs is either consumed by another agent or selected by an evaluator. Credit is explicitly not money; it is a local reinforcement signal that the agent’s transformation rule has acquired a functional role in the network. If an agent repeatedly receives no credit, its probability of adaptation rises with the number $\Delta t$ of consecutive uncredited episodes, and local mutation perturbs $a_i$, $P_{dis}$0, and $P_{dis}$1 with $P_{dis}$2, after which $P_{dis}$3 is renormalized (Metzner et al., 25 Jun 2026).

This mechanism makes SBS neither a standard global optimizer nor a purely static network model. It is a dynamics in which stabilization depends on downstream uptake, while lack of uptake keeps a component plastic.

3. Emergent organization, novelty, and pre-adaptive search

The principal empirical claim of the SBS paper is that these local rules generate organized structure. In a small system with $P_{dis}$4 and $P_{dis}$5, one agent can act immediately on the raw materials; the paper reports that agent 4 accepts both raw materials in episode 1 and that its output is already accepted by evaluator 0 in the same episode, while the stricter full-target criterion is reached only after about 10 more episodes. The eventual successful state for evaluator 1 is not a direct raw-material combination but arises from a short transformation chain involving other agents, demonstrating that SBS can generate multi-step functional pathways rather than only one-shot outputs (Metzner et al., 25 Jun 2026).

A more stringent test is the Missing Dimension Hurdle (MDH). In that setup, both raw materials have zero amplitude in feature component 0, while both evaluators depend exclusively on that component, with targets $P_{dis}$6 and $P_{dis}$7. Because linear combinations of the raw inputs cannot leave the initial subspace, success depends on the out-of-subspace term in the production rule. With $P_{dis}$8, the system initially hovers near the $P_{dis}$9 axis, novelty remains tiny during roughly the first 30 episodes, and the first full-target condition appears around $N_{dim}$0. The authors interpret this as a pre-adaptive search phase: the system explores and accumulates usable intermediates before external success is even possible (Metzner et al., 25 Jun 2026).

At larger scale, SBS yields a core-periphery organization. In a system with $N_{dim}$1 under MDH, evaluator activity is disabled for the first 500 episodes, and transfer matrices $N_{dim}$2 are computed over 50-episode windows, where $N_{dim}$3 counts how often a state generated by agent $N_{dim}$4 is used as input by agent $N_{dim}$5. The resulting networks show a stable core of active agents—10 agents in one run and 6 in another—surrounded by a dynamic periphery. Core agents exhibit dense mutual utilization, regular credit, and almost no adaptation; peripheral agents exhibit sparse interactions, low credit, and ongoing adaptation. Enabling evaluators later does not significantly disrupt the internal core structure (Metzner et al., 25 Jun 2026).

The threshold sweep further constrains interpretation. As the acceptance threshold is swept from 0 to 1 over 50 repetitions per value, the probability of satisfying the full-target criterion, the number of evaluator selections per episode, and agent-to-agent transfers all decrease with threshold, while adaptation increases. Interaction-space size and state diversity show non-monotonic peaks at intermediate thresholds. This suggests that SBS is most productive when compatibility is neither too permissive nor too restrictive (Metzner et al., 25 Jun 2026).

4. SBS-equivalent formulations in self-training

A closely related formulation appears in “Survival is the Only Reward: Sustainable Self-Training Through Environment-Mediated Selection” (Dodgson et al., 18 Jan 2026). That work does not officially name its framework SBS, but it explicitly presents an equivalent selection principle: candidate behaviors are executed under real resource constraints, and only those whose environmental effects both persist and preserve the possibility of future interaction are propagated. The environment provides no semantic feedback, dense rewards, task-specific supervision, or human-curated gold dataset as the primary driver (Dodgson et al., 18 Jan 2026).

In the proof-of-concept implementation, survival is operationalized as a conserved physical resource, specifically non-volatile memory occupancy or available storage space. The environment is a procedurally generated containerized Linux setting with realistic filesystem structures, permissions, processes, and passwords. The model proposes executable code, the code is run in the sandbox, the environment changes, and only behaviors associated with sum-positive environmental impact are retained for training. The paper formalizes the interaction as

$N_{dim}$6

with each trajectory $N_{dim}$7 inducing a measurable environmental consequence $N_{dim}$8, and the selection rule

$N_{dim}$9

In this sense, behaviors survive by serving the environmental survival constraint rather than by optimizing a proxy reward (Dodgson et al., 18 Jan 2026).

The paper’s conceptual term for the resulting learning dynamic is negative-space learning (NSL). Learning proceeds through consolidation and pruning: the system does not primarily accumulate a large positive catalog of behaviors, but reallocates probability mass toward a smaller set of strategies that continue to work. The history of successful trajectories is

$N_{dim}=5$0

and the online update is written generically as

$N_{dim}=5$1

where $N_{dim}=5$2 is the replay buffer of recent successful trajectories. Two memory-bounded regimes are contrasted. The Miri regime uses only the last three successful runs, enforcing temporal locality; the Katalin regime uses the top three trajectories by $N_{dim}=5$3 from the entire history. The authors report that Miri is more stable, whereas Katalin can exhibit gradient conflict and collapse (Dodgson et al., 18 Jan 2026).

The same paper reports an emergent meta-learning phenomenon described as deliberate experimental failure: pass@1 on code generation can drop to zero in some later online generations even while broader code quality improves, which the authors interpret as the model learning to use failure and error messages instrumentally. This suggests an SBS-compatible interpretation in which some behaviors serve not immediate return but the acquisition of informative environmental feedback (Dodgson et al., 18 Jan 2026).

5. Engineering SBS in networked systems and operating systems

In SDN/NFV, SBS is instantiated as proactive service-chain survivability. An NFV infrastructure is modeled as a graph $N_{dim}=5$4, where each physical node $N_{dim}=5$5 has capacity $N_{dim}=5$6, and the infrastructure hosts many service function chains composed of VNFs of type $N_{dim}=5$7. Because VNFs from different chains can be co-located on the same physical node, failure of a single node can knock out multiple VNFs and thereby break several SFCs at once. The proposed survivability management framework contains a Service Chain Provisioning Module, Monitoring Module, Rerouting Module, and Backup Provisioning Module; the core contribution is the proactive provisioning of shared backup VNFs before any failure occurs (Aidi et al., 2018).

The backup problem is formulated as an ILP with decision variables $N_{dim}=5$8 for the number of type-$N_{dim}=5$9 backups on node $S(\vec{X},\vec{Q})=\vec{X}\cdot\vec{Q},$0 and $S(\vec{X},\vec{Q})=\vec{X}\cdot\vec{Q},$1 for the assignment of VNFs of type $S(\vec{X},\vec{Q})=\vec{X}\cdot\vec{Q},$2 on node $S(\vec{X},\vec{Q})=\vec{X}\cdot\vec{Q},$3 to backup host $S(\vec{X},\vec{Q})=\vec{X}\cdot\vec{Q},$4. The objective is

$S(\vec{X},\vec{Q})=\vec{X}\cdot\vec{Q},$5

subject to constraints that prohibit hosting a backup on the same node as the protected VNF, require exactly one backup host assignment per protected set, enforce capacity, and bound synchronization distance by $S(\vec{X},\vec{Q})=\vec{X}\cdot\vec{Q},$6. Two heuristics, BS-Pull and BS-Push, provide scalable approximations. In simulations on a 24-node, 55-link infrastructure with node capacities randomly generated from 20 to 50 VMs and $S(\vec{X},\vec{Q})=\vec{X}\cdot\vec{Q},$7, the heuristics are near-optimal at low utilization, BS-Pull generally uses fewer backups than BS-Push, and both methods maintain nearly flat execution time while CPLEX increases from about 2 seconds up to around 7 minutes in scenario S6. Under very high utilization, full backup coverage becomes impossible because the infrastructure lacks spare capacity (Aidi et al., 2018).

In operating-system security, the same intuition appears as survivability through constrained continuation of service. “Survivor: A Fine-Grained Intrusion Response and Recovery Approach for Commodity Operating Systems” proposes an orchestration of fine-grained recovery and per-service responses after intrusion detection. The architecture combines an IDS, a response-selection component, a service manager extended from systemd, and storage and logging infrastructure protected by kernel-enforced isolation using MAC/SELinux-like controls. During normal operation, the service manager checkpoints services, snapshots the filesystem, and records which files each monitored service modified. After an alert, the system restores the service to the last known safe checkpoint, restores only the files modified by that service, applies selected per-service restrictions, and resumes service in a possibly degraded mode (Chevalier et al., 2019).

The response-selection procedure is explicitly cost-sensitive. For each malicious behavior $S(\vec{X},\vec{Q})=\vec{X}\cdot\vec{Q},$8, the system filters candidate responses $S(\vec{X},\vec{Q})=\vec{X}\cdot\vec{Q},$9 by excluding those whose cost is critical, computes a Pareto-optimal set, and selects a response by maximizing

$S(\vec{X},\vec{Q}) \ge \Theta_{thr}$0

with weights derived from qualitative risk. The qualitative scale runs from none through very low, low, moderate, high, very high, to critical. The operative SBS logic is that non-essential or dangerous functionality is sacrificed so that core functions remain available. In the experiments, the prototype added about 3561 lines of C code across CRIU, systemd, Linux audit user space, and kernel audit-side modifications; checkpoint time was typically under 300 ms, restore time typically under 325 ms, and overhead was generally small except for the rare case of services that write many small files asynchronously in a burst (Chevalier et al., 2019).

6. Relations to adjacent theories, misconceptions, and limits

SBS is often misconstrued as generic survival maximization. The literature summarized here is narrower. In the canonical self-organization paper, persistence depends specifically on downstream utilization rather than on raw longevity or externally defined reward (Metzner et al., 25 Jun 2026). In the self-training paper, the authors explicitly contrast environment-mediated selection with reinforcement learning, emphasizing that the method uses no explicit reward model, no semantic feedback, no dense shaping, and no task-specific supervision, and that simple supervised fine-tuning on successful trajectories was more stable than PPO, GRPO, or DPO in their proof of concept (Dodgson et al., 18 Jan 2026). In the engineering papers, the criterion is still not “survive at any cost”; it is maintain service continuity subject to capacity, synchronization, or policy constraints (Aidi et al., 2018, Chevalier et al., 2019).

The concept also overlaps with, but is not identical to, several established frameworks. The SBS self-organization paper compares its dynamics with hypercycle theory, autocatalytic networks, RAF sets, autopoiesis, closure of constraints, economic utilization, reservoir computing, SORN-like self-organizing recurrent networks, backpropagation, and modular neural architectures. Its claimed distinction is that persistence depends on continued utilization rather than only on network closure or global supervised error (Metzner et al., 25 Jun 2026). The self-training paper similarly notes an evolutionary analogy—behaviors are tested, unsuccessful variants disappear, and successful ones are retained and recombined—but emphasizes that the selector is environment-grounded persistence rather than a formal proxy fitness function (Dodgson et al., 18 Jan 2026).

A second misconception is terminological overextension. Not every “survival” problem is an SBS problem. “Optimal Surviving Strategy for Drifted Brownian Motions with Absorption” studies Aldous’s “Up the River” problem, where a unit drift budget is allocated among absorbed Brownian particles and the asymptotically optimal policy is push-the-laggard, yielding approximately $S(\vec{X},\vec{Q}) \ge \Theta_{thr}$1 surviving particles as $S(\vec{X},\vec{Q}) \ge \Theta_{thr}$2 (Tang et al., 2015). That result concerns optimal control of survival under absorption, not persistence through serving downstream utilization.

The main limitations are domain-specific. The canonical SBS model is highly simplified: agent populations are fixed, transformation rules are simple, target conditions are externally imposed, there is no explicit memory or goal-directed planning, and the credit scheme may stabilize internally coherent cores that are only weakly coupled to external relevance (Metzner et al., 25 Jun 2026). The self-training instantiation presents proof-of-concept evidence rather than a formal proof that reward hacking is impossible in all settings, and it reports instability under the Katalin regime (Dodgson et al., 18 Jan 2026). The NFV formulation assumes single-node failures, same-type backups, and bounded synchronization distance, and it can fail to protect all VNFs when spare capacity is insufficient (Aidi et al., 2018). The OS-security approach depends on reliable intrusion detection and policy specification, and accepts degraded service as the survivability state rather than full functional restoration (Chevalier et al., 2019).

Taken together, these works support a precise but plural understanding of SBS. At its most general, SBS is a substrate-independent proposal that functional utilization can serve as a distributed selector of organized structure (Metzner et al., 25 Jun 2026). In adjacent machine-learning and systems literatures, the same logic reappears as survival through environment-grounded retention, proactive redundancy, or restricted continuation of core service (Dodgson et al., 18 Jan 2026, Aidi et al., 2018, Chevalier et al., 2019). The unifying theme is not mere persistence, but persistence conditional on remaining functionally relevant.