Paper deep dive
Jevolution: A Framework for Studying Communication and Collective Dynamics in AI Agents
Xirui Huang, Nikhil Hooda, Utkarsh Gupta
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 10/14/2026, 9:57:47 AM
Summary
The paper introduces Jevolution, an instrumented spatial predator-prey framework designed to study communication and collective dynamics in AI agents. It addresses the complexities of asynchronous decision-making, world evolution during inference, and demographic changes. The framework separates the evolving world, scheduling, controller adapters, and experiment records, utilizing the ODD protocol for specification. A case study on rabbit and wolf populations with signal-based communication showed negligible effects on abundance, emphasizing the tool's role as an inspectable experimental system rather than a demonstration of emergent intelligence or biological validity.
Entities (8)
Relation Signals (7)
Jevolution → simulates → Rabbits
confidence 95% · Rabbits and wolves act on a spatial habitat... Jevolution provides an artificial environment...
Jevolution → simulates → Wolves
confidence 95% · Rabbits and wolves act on a spatial habitat... Jevolution provides an artificial environment...
Jevolution → uses → ODD Protocol
confidence 95% · We describe these mechanisms using the ODD protocol...
Jevolution → affiliatedwith → University of Waterloo
confidence 90% · Author affiliations listed in the paper header.
Jevolution → createdby → Xirui Huang
confidence 90% · Author list and abstract presentation of the framework.
Jevolution → contrastswith → Lotka-Volterra
confidence 85% · We do not fit Lotka-Volterra coefficients or infer those equations from visually cycling trajectories.
Jevolution → usescontrollerfrom → TypeSafe AI
confidence 80% · In the evaluated adapter, Jev jev-1.13.0 answers... (TypeSafe AI, 2026b...)
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Experiments with remote decision-making agents depend on more than the policy: the world may change during inference, the simulator determines which actions are available, and births and immigration alter who can act. We present Jevolution, an instrumented spatial predator-prey framework that makes these dependencies explicit. Its reference implementation separates the evolving world, asynchronous request scheduling, controller adapters, and experiment records. The model combines local observations, energy and resource constraints, predefined signals, and inherited scalar traits. We describe these mechanisms using the ODD protocol and distinguish implemented controls from research questions requiring additional interventions. Free synthetic fixtures exercise scheduling, signaling, and inheritance against the pinned implementation; they validate specified software behavior rather than model intelligence or ecological realism. A completed communication case study illustrates the end-to-end workflow using Jev to control both species. Across 6 complete habitat-seed pairs, the mean signals-on minus signals-off difference in normalized rabbit abundance was -0.0095 (nominal 95% pair-bootstrap interval [-0.0772, 0.0602]). This small study does not resolve the direction of the signal-system effect. The contribution is an inspectable experimental system and its evidence boundaries, not a new policy-learning algorithm or a demonstration of emergent language, adaptive selection, or biological predictive validity.
Tags
Links
- Source: https://arxiv.org/abs/2610.14449v1
- Canonical: https://arxiv.org/abs/2610.14449v1
Trouble viewing inline? Open PDF directly →
Full Text
95,808 characters extracted from source content.
Expand or collapse full text
Jevolution: A Framework for Studying Communication and Collective Dynamics in AI Agents Xirui Huang Nikhil Hooda Utkarsh Gupta University of Waterloo University of Waterloo University of Waterloo x36huang@uwaterloo.ca n7hooda@uwaterloo.ca u2gupta@uwaterloo.ca Abstract Experiments with remote decision-making agents depend on more than the policy: the world may change during inference, the simulator determines which actions are available, and births and immigration alter who can act. We present Jevolution, an instrumented spatial predator–prey framework that makes these dependencies explicit. Its reference implementation separates the evolving world, asynchronous request scheduling, controller adapters, and experiment records. The model combines local observations, energy and resource constraints, predefined signals, and inherited scalar traits. We describe these mechanisms using the ODD protocol and distinguish implemented controls from research questions requiring additional interventions. Free synthetic fixtures exercise scheduling, signaling, and inheritance against the pinned implementation; they validate specified software behavior rather than model intelligence or ecological realism. A completed communication case study illustrates the end-to-end workflow using Jev to control both species. Across 6 complete habitat-seed pairs, the mean signals-on minus signals-off difference in normalized rabbit abundance was -0.0095 (nominal 95% pair-bootstrap interval [-0.0772, 0.0602]). This small study does not resolve the direction of the signal-system effect. The contribution is an inspectable experimental system and its evidence boundaries, not a new policy-learning algorithm or a demonstration of emergent language, adaptive selection, or biological predictive validity. Preprint. Revised 1 October 2026. Project website | Source repository 1 Introduction A simulation can change while an agent is deciding what to do. A resource may disappear, a target may move, and a warning may expire before a response arrives. In a population of agents, the rate at which decisions become available also depends on who is alive, who is eligible to act, and how requests share finite concurrency. These details can change the meaning of a policy comparison. Real-time and delayed reinforcement learning provide precedents for treating timing as part of the interaction process (Ramstedt and Pal, 2019; Ramstedt et al., 2020). Jevolution provides an artificial environment in which these dependencies can be inspected together. Rabbits and wolves act on a spatial habitat with food, water, terrain, predation, reproduction, and immigration. A controller chooses from engine-generated actions while the world continues to advance. Rabbit signals and inherited traits provide mechanisms for posing questions about communication and population change. They do not, merely by existing, establish emergent coordination or natural selection. Our central systems question is: what must be exposed and recorded to make experiments with asynchronous, interacting populations interpretable? We answer through a concrete implementation, explicit process semantics, executable engineering checks, and a completed case study. The reference task is predator–prey; we do not claim a general simulator for arbitrary social or biological systems. The intended user is a researcher who needs to inspect observation construction, schedule policy calls, modify a mechanism, and trace a population outcome back to a reproducible run record. The paper contributes: (1) an architecture and ODD specification connecting world dynamics, controller choices, scheduling, and records (Sections 3–4); (2) an evidence map for timing, communication, emergence, and inheritance, separating implemented mechanisms from unresolved hypotheses (Section 5); and (3) software-level validation plus a paired communication case study demonstrating the experimental workflow (Section 6). The source and analysis are pinned locally. Figure 2 shows the interactive interface; Appendix A illustrates recorded experimental states. The contribution is this concrete integration and its inspectability; no superiority over existing platforms, new learning algorithm, or independently adopted benchmark is established. 2 Related work LLM agents and simulation. Generative Agents combines memory, reflection, and planning in an interactive social world (Park et al., 2023). Project Sid describes a real-time architecture for populations of agents in Minecraft (Altera.AL et al., 2024), while AgentSociety integrates agents, an environment, and a simulation engine for social interventions (Piao et al., 2025). Affordable Generative Agents addresses the cost of repeated inference through reusable behavior and compressed context (Yu et al., 2024). A survey of LLM-based simulation distinguishes agent construction, alignment, and evaluation challenges (Gao et al., 2024). These works motivate careful agent and runtime descriptions. Their results do not supply external validity for Jevolution’s animal-like agents. Agent-based modeling and testbeds. NetLogo and Mesa already provide environments and components for constructing agent-based simulations (Wilensky, 1999; Masad and Kazil, 2015). The ODD protocol and its updates organize model purpose, entities, scheduling, initialization, inputs, and submodels to support understanding and reimplementation (Grimm et al., 2006; Grimm et al., 2010; Grimm et al., 2020). The flame framework, linked to the AgentTorch implementation, supports differentiable agent-based models for calibration and action learning (Chopra et al., 2024). Jevolution does not implement that learning method. MAgent and Neural MMO supply many-agent settings, including spatial interaction and competition for resources (Zheng et al., 2018; Suarez et al., 2019). Melting Pot emphasizes evaluation with unfamiliar partners and social situations (Leibo et al., 2021); PettingZoo makes multi-agent interaction order explicit through a standard API and agent-environment-cycle model (Terry et al., 2021). We make no claim to exceed these platforms in scale or evaluation breadth. Communication and adaptation. RIAL/DIAL, CommNet, and learned referential games optimize communication along with agent behavior (Foerster et al., 2016; Sukhbaatar et al., 2016; Lazaridou et al., 2016). Grounded multi-agent communication has also been studied in physical environments (Mordatch and Abbeel, 2018). Jevolution instead exposes a predefined vocabulary to a remotely served policy. Message use is therefore distinct from discovering a new language. Following Lowe et al. (2019), a correlation between a signal and an action is insufficient evidence of receiver influence. In biological communication, playback of different vervet alarm calls elicited different defensive responses without a predator being present (Seyfarth et al., 1980); this illustrates an intervention on received information, not an equivalence between text symbols and natural signals. Similarly, hide-and-seek autocurricula train policies (Baker et al., 2020), and Evolutionary Policy Optimization combines population search with policy gradients (Wang et al., 2025). Neither is the same operation as mutating an offspring’s simulation traits while leaving policy training outside the experiment. Ecological context. Lotka and Volterra analyze coupled populations and cyclic solutions under explicit assumptions (Lotka, 1925; Volterra, 1926). Controlled population experiments by Gause and spatial predator–prey experiments by Huffaker expose the importance of experimental conditions, dispersion, and access to prey (Gause, 1934; Huffaker, 1958). Rosenzweig and MacArthur show how interaction structure affects stability (Rosenzweig and MacArthur, 1963). These sources motivate explicit ecological mechanisms and restrained interpretation. Jevolution’s immigration, terrain, capacity limits, and decision scheduling prevent inferring a Lotka–Volterra mechanism from the appearance of its population traces. Timing and evaluation. Real-Time Reinforcement Learning models an environment that evolves during action selection (Ramstedt and Pal, 2019); random-delay formulations separately describe observation and action delays (Ramstedt et al., 2020). Thus, reporting only provider response time omits other contributors to stale actions. Work on reproducibility and uncertainty in deep-RL evaluation also warns against treating a few stochastic runs as decisive evidence (Henderson et al., 2018; Agarwal et al., 2021). We use these as methodological precedents rather than as evidence that this experiment trains a deep-RL policy. The distinction in this paper is the specified combination of asynchronous controller calls, demographic turnover, engine-defined interaction mechanisms, and request-linked records in one pinned task. Appendix B compares documented emphases without inferring absent features or measured superiority. The contribution is an executable reference for these experimental concerns; comparative practical advantage remains untested. 3 Architecture and experimental workflow World, runtime, controller, and record. Here, a runtime lineage means a configured controller group, not an inferred genealogical lineage. Individual parent IDs and generations are separate ancestry fields. The reference architecture has four responsibilities (Figure 1). The world stores spatial and demographic state and applies physics. The runtime advances that world, identifies eligible actors, and manages outstanding requests. A controller adapter receives a local observation and legal choices and returns a structured decision. Experiment hooks record observations, outcomes, state summaries, and provenance. These boundaries allow a synthetic controller to exercise scheduling without contacting a provider, and allow recorded executions to be analyzed without rerunning inference. Figure 1: Architecture of the pinned implementation. Arrows show information flow, not synchronous execution; world physics continues during requests. The controller receives local observations and engine-offered choices; the engine resolves routes and applies effects. Request identifiers join records across both paths. This schematic is not an empirical result, so no seed count or confidence interval applies. Inspecting the world and its decisions. The interface brings three views of a run together: population history, the spatial world, and a selected animal’s state and decision (Figure 2). This makes it possible to inspect a local decision alongside the world in which it occurs. The screenshot illustrates these views in the historical prototype; the displayed deployment has not been matched to the frozen experimental revision. For the case study, analysis uses the retained experiment records. Appendix A visualizes saved states from one experimental pair. Figure 2: Jevolution’s interactive interface. Left: run controls and population history. Center: habitat, animals, and signal markers. Right: the selected animal’s decision and state inspector. The image is reproduced unchanged from Figure 1 of the authors’ original manuscript, at 31.59 simulation seconds; it is not one of the twelve 90-s study runs. Visible signals illustrate the interface, not evidence of a communication benefit. Interface and artwork provenance are documented in Section 8 and Appendix K. Two clocks and three delays. Simulation time describes the world; wall time controls request acceptance and service waiting. The reference runtime integrates elapsed time using physics substeps no longer than 50 simulation ms. In the live case study, simulation speed is 1×1×, a returned decision must pass a 2,500-wall-ms acceptance threshold, and at most six requests per lineage may be outstanding. A default 100-simulation-ms post-request cooldown has engine-triggered eligibility resets; it is not an unconditional per-animal minimum. An additional global dispatch spacing of 100 wall ms bounds the case-study request rate. Transport timeout, response acceptance, and scheduling eligibility are distinct mechanisms. Let tet_e, tdt_d, and tat_a denote eligibility, observation/dispatch, and application times on the simulation clock. Queue delay is td−tet_d-t_e; applied observation age is ta−tdt_a-t_d. Separately, provider-adapter elapsed time is measured on the wall clock and includes transport and local handling. A long eligibility queue is not necessarily a stale observation because the observation is captured at dispatch. These definitions prevent treating every delay as model compute or every waiting animal as holding an old decision. Controller boundary. The controller receives local state and engine-generated legal choices, not unrestricted control over physics. Route generation may inspect the map beyond the observation; therefore feasible navigation is partly supplied by the engine. In the evaluated adapter, Jev jev-1.13.0 answers independent action and signal Choice questions sharing one observation (TypeSafe AI, 2026b; TypeSafe AI, 2026c). Neither answer conditions on the other. The code combines both choices into one simulator decision; returned confidence and per-option probabilities are logged but not used for selection or gating. The provider describes this version as a structured decision model without text-generation training (TypeSafe AI, 2026a). Its internal architecture and weights were not independently inspected. From interactive prototype to experiment. A reproducible run specifies the source revision, effective rules, initial world, controller identity, clocks, concurrency, dispatch policy, horizon, intervention, and stopping rules. Source-defined controls must be distinguished from interface previews: a visible control alone does not establish a changed mechanism. In the pinned study implementation, communication-off is enforced in the engine. Request, response, and application/error records share identifiers; population counts are retained after each physics substep, with complete initial/final states and approximately one-second world snapshots. The bounded interface journal is not substituted for this append-only experiment record. The Node host and browser worker both instantiate this core runtime with different persistence and decision-service implementations. This is shared source, not a demonstrated claim of identical wall-clock behavior across hosts. A type-checked local-controller example and the limits of the virtual-clock harness appear in Appendix C. These choices support three operations: changing a mechanism with a source-level test, swapping a controller at the decision boundary, and reanalyzing recorded outcomes. Adding a new species, training algorithm, or ecological model still requires implementation work. Configurability is not evidence that every combination is scientifically meaningful or has been tested. The artifact’s reuse contract is a pinned, executable research implementation, not a stable general-purpose plugin API. Configuration, service substitution, and source changes must be distinguished (Table 1). Table 1: Actual intervention routes on the frozen source. This is a source-backed capability map, not a usability experiment. Route Supported operation and boundary Run configuration Seed, communication condition, timing settings, and audit runtime options. Predator–prey reset installs a fixed Jev roster: supplied provider, population, and group-delay roster edits are replaced. Arena roster configuration is a separate path. Injected service A host may replace choose while keeping the engine and observation builder. Synthetic responses test this boundary; virtual clocks additionally require the white-box harness described in Appendix C. Source changes New species, a supported clock port, receiver-only/fixed-trait controls, or policy optimization need additional implementation and validation. Frozen UI preview metadata is not an active experimental control. 4 Reference model in ODD form We organize the reference model by the seven ODD elements (Grimm et al., 2006; Grimm et al., 2010; Grimm et al., 2020); Appendix D expands the description. The implementation is based on commit 9285979e95d67d8f11fa4f6d9b603b156c84380 plus the frozen experimental patch identified in Appendix F. Purpose and patterns. The task exposes interactions between local information, asynchronous decisions, and population outcomes. It is not fitted to a real ecosystem. Clustering or oscillating counts are observations to explain, not automatic validation criteria. Entities, state, and scales. A 64×6464× 64 habitat contains terrain and food. Animals have identity, position, energy, age, lineage, activity, and scheduling state. Rabbits also have bounded memory and five scalar traits: speed, vigilance, thrift, fertility, and sociability. Signals have a source, position, type, delivery time, and expiry. The reference predator–prey initialization contains 70 rabbits and 8 wolves; effective parameters are retained per run. Process overview. Physics updates resource availability, movement, metabolism, predation, reproduction, and immigration while requests are outstanding. Substep order is specified in Appendix I. Newborn insertion and the ordering of rabbit and wolf updates can affect trajectories; asynchronous returned choices follow a separate runtime path. Design concepts. Policies operate with partial observations, predefined signals, and legal actions. Both species reproduce automatically and from a single parent in this preset. The remotely served controller is not trained inside a run. Inherited simulator traits affect mechanics; they are not inherited model weights. Engine rules, resource geometry, and birth placement can produce spatial structure without message use. Initialization and inputs. Habitat seeds produce recorded initial states. Paired conditions match biological-state hashes; the full configuration hash includes the intervention. Later trajectories can diverge through choices, completion order, and state-dependent random draws. Input data are synthetic world state and controller responses, not measurements from animals. The source is the implementation specification; saved effective configurations prevent an unversioned preset name from standing in for actual settings. Submodels. For sociability s, signal radius is 3+5s3+5s and immediate energy cost is 0.6(1+s)0.6(1+s). Delivery is scheduled 0.2 simulation seconds after emission and expiry six seconds after emission. The global history retains at most 200 signals; oldest-entry eviction can end availability earlier than the nominal expiry. Terrain, range, and observation timing restrict exposure. The predator–prey vocabulary is danger, food, follow, and silence; help and wolf scent signaling are disabled in this preset. Child rabbit traits follow gchild=clip[0.05,0.95](gparent+ϵ),ϵ∼[−0.065,0.065),g_ child=clip_[0.05,0.95](g_ parent+ε), ε [-0.065,0.065), (1) with a finite pseudorandom implementation. Food uses capped linear regrowth, not a logistic population equation. Capacity, energy, age, placement, and cooldown constrain births. Immigration attempts add two rabbits and one wolf after at least 20 simulation seconds since the prior eligible check, subject to placement. Between resets and external edits, counts obey N(t)=N(0)+B(t)+I(t)−D(t).N(t)=N(0)+B(t)+I(t)-D(t). (2) Births B and immigrants I must remain distinct: temporary zero abundance need not be permanent extinction, and abundance is not individual fitness. 5 Research scope and evidence requirements The framework supports a research agenda, but each question needs its own identifying intervention. Table 2 separates current mechanisms and evidence from unresolved questions. A broad framework contribution does not turn these questions into completed findings. Table 2: Research scope of the reference system. These are capability/evidence distinctions, not empirical estimates; no seed count or interval applies. The case study and synthetic checks are reported separately. Topic Implemented basis Evidence still needed for a scientific claim Inference latency Asynchronous requests, eligibility, deadlines, explicit clocks A controlled delay intervention with matched policy, observation, and scheduling; population effects are not established by deadline tests. Communication Predefined messages; engine-enforced on/off mechanism Receiver-only and cost-matched interventions to separate information use from energy/option changes. Emergent behavior Interacting agents, spatial resources, births, immigration Repeated patterns and spatially matched nulls preserving resource attraction and demographic placement. Natural selection Inherited traits, ancestry state, and immigration counters; cohort reconstruction required Cohort/lineage reproductive success and fixed-trait or inheritance controls separating mutation, drift, immigration, and selection. Policy learning Controller adapter boundary A training loop, learning objective, and evaluation protocol; EPO is not implemented. Timing as an intervention. Injecting a known response delay can verify deadline handling under a controlled clock. To infer the causal effect of latency on collective outcomes, however, an experiment must also specify whether observations are captured before or after the delay, how pending calls consume concurrency, and whether the world advances during waiting. A synthetic deadline fixture answers the software question; it is not a population-effect experiment or a measurement of the provider’s inference latency. Communication and emergence. Message production, exposure, and influence are different observations. A predefined warning alphabet does not establish learned language. Receiver interventions are needed to assess influence (Lowe et al., 2019). Likewise, a cluster near food can be generated by shared attraction, births near parents, or immigration beside conspecifics. A proposed emergence study would compare nearest-neighbor or recruitment statistics against spatially matched nulls retaining those mechanisms, and repeat the comparison over run-level replicates. Such null analyses have not been performed here. Inheritance, natural selection, and EPO. Equation 1 supplies variation and inheritance at the simulator-trait level. A changing average trait is insufficient to attribute adaptive change when arrivals, mutation, mortality, and reproduction all contribute. Genealogical outcomes and interventions are required. A controller-group label is not an ancestry tree; rabbit single-parent births store the same parent ID twice, so parent counts must be deduplicated when defining ancestry. Evolutionary Policy Optimization instead combines population search with policy gradients (Wang et al., 2025); Jevolution has no such policy-optimization loop. A future implementation would need explicit policy identities, fitness/evaluation rules, and training data beyond the current controller adapter. Ecological reference rather than ecological equivalence. The classical Lotka–Volterra equations describe aggregate populations under explicit assumptions (Lotka, 1925; Volterra, 1926). Jevolution instead applies spatial, resource, birth, immigration, and decision rules. We do not fit Lotka–Volterra coefficients or infer those equations from visually cycling trajectories. Gause, Huffaker, and Rosenzweig–MacArthur motivate attention to experimental conditions and interaction structure (Gause, 1934; Huffaker, 1958; Rosenzweig and MacArthur, 1963). Ecological predictive use would require external measurements, calibration, and independent validation. 6 Validation and worked example 6.1 Software behavior under controlled inputs A separately declared free protocol tests the pinned source with deterministic injected responses and a virtual clock. It holds a small artificial world away from ecological interactions, crosses six response delays with two per-lineage capacities and two seeds, and repeats every condition exactly. This yields 48 executions of 24 conditions, plus 4 executions of pause/removal fixtures. The deadline is deliberately set to 1,000 virtual wall ms; these fixtures are not the live Jev configuration. All matrix executions reach the exact 4.137-s horizon, and pending calls at the horizon remain recorded as cancellations. Observation and terminal records join by request ID. Responses at the deadline apply, whereas those one millisecond beyond it are late (Table 3). At a fixed 100-ms response delay, capacity one yielded 74 applied decisions per execution and capacity three yielded 120: the response time alone did not determine decision opportunity in this fixture. These counts are not a live-service throughput benchmark. The outstanding-call caps hold in the tested matrix; every dispatched request has one terminal outcome. Repeated runs match the defined canonical state/trace hash. This is schedule-conditioned reproducibility under controlled inputs, not bitwise repeatability of a remote service or a proof for untested schedules. Table 3: Synthetic runtime fixtures, not provider throughput: request/applied/late/cancelled counts per execution over 4.137 simulation seconds. Each cell was run for two seeds (42, 314159), each repeated twice; all four executions in a cell had the shown counts. There is no sampling-based interval for these deterministic fixtures. Capacity is per lineage, with three animals in each of two lineages. The acceptance deadline is 1,000 virtual wall ms, distinct from the live study’s 2,500 wall ms. Injected delay (ms) Capacity 1 Capacity 3 0 238 / 238 / 0 / 0 240 / 240 / 0 / 0 100 76 / 74 / 0 / 2 126 / 120 / 0 / 6 500 18 / 16 / 0 / 2 42 / 36 / 0 / 6 1000 10 / 8 / 0 / 2 24 / 18 / 0 / 6 1001 10 / 0 / 8 / 2 24 / 0 / 18 / 6 1500 6 / 0 / 4 / 2 18 / 0 / 12 / 6 Separate fixtures cover 72 signal boundary cells, 96 reproduction conditions (each repeated), and 16 empty-population immigration conditions (each repeated). They verify the tested delivery/range/expiry rules, mutation bounds and ancestry, and arrival accounting. All 2,724 recorded invariant checks pass in the completed suite. One earlier harness attempt stalled on a terminal floating-point clock boundary; its partial records and the harness correction are retained. No engine source changed. Appendix E gives the fixtures and provenance. These results establish specific software behavior, not adaptation, model competence, external usability, or ecological validity. 6.2 Communication as an end-to-end case study The worked example asks whether enabling the implemented rabbit signal system changes population outcomes with Jev controlling both species. It reuses the completed, locally preregistered study; it is not new data collected for this framing. Each of six habitat seeds has one signals-on and one signals-off run, with identical initial biological-state hashes and a fixed 90-s horizon. All twelve main runs reached that horizon; no pair was excluded or replaced. The order was randomized once, yielding five on-first pairs and one off-first pair. No drought was applied. Appendix F retains all seeds, order, source, and stopping rules. Signals-off blocks non-silent emission, reception, signal-derived choices, and signal energy charges while preserving personal visual memory and visually supported following. Thus the intervention changes the entire signal system, not solely the information value of messages. The endpoint is rabbit population area under the curve divided by 90×7090× 70, using left-held post-update counts from every engine substep. It includes descendants and immigrants and is not individual fitness. The estimator is the mean within-seed on-minus-off difference. The nominal interval resamples whole seed pairs 10,000 times with analysis seed 20260928 (Efron, 1979). Animals and requests are not independent replicates. Across 6 complete habitat-seed pairs, the mean signals-on minus signals-off difference in normalized rabbit abundance was -0.0095 (nominal 95% pair-bootstrap interval [-0.0772, 0.0602]). This small study does not resolve the direction of the signal-system effect. Figure 3: Communication worked example: all n=6n=6 paired normalized rabbit-AUC effects and their mean. Each pair contains one 90-s run per arm. The nominal 95% percentile-bootstrap interval uses 10,000 whole-pair resamples (seed 20260928); few-pair coverage is unreliable. These are the same recorded runs as the communication-focused draft, not an independent replication. The leave-one-pair-out means range from -0.0379 to 0.0124. This sensitivity and the interval spanning zero limit the case-study conclusion. The contribution to the framework argument is the traceable route from configuration through requests and mechanism intervention to a qualified run-level estimate, not a positive communication result. Operational evidence and cost. Main runs submitted 10,514 requests, of which 10,462 were applied. The full inventory retains late, invalid, error, and cancelled outcomes, including transport failures, rather than dropping their runs. Across all paid attempts, including pilots and unsuccessful calls, returned usage corresponds to $1.226087; 24 requests without settled usage retain $0.066060 in full-context reservations, giving $1.292148 in conservative task accounting. No additional paid inference was used for the framework validation. These are token-derived charges and bounded liabilities, not an independently reconciled per-run invoice. The shared attempt inventory retains the two pilots, two engineering smoke runs, twelve free-controller runs, and twelve main runs. The free controller ignores received messages and has zero provider latency; it is a separate descriptive reference, not a matched Jev comparison. The retained timing records distinguish eligibility waiting from observation age and provider-adapter elapsed time (Appendix J). On-condition rabbit snapshots expose received messages in 828 of 2,634 observations; off-condition snapshots do so in zero of 2,601. These are exposure observations, not distinct deliveries or evidence of listening. The case study establishes neither learned language nor adaptive selection, a causal population effect of latency, or ecological predictive validity. 7 Limitations and intended use The framework has one fully instrumented, remotely controlled case study in one preset. The synthetic fixtures test selected implementation contracts; they do not measure biological validity, service performance, scalability, developer productivity, or usefulness to independent users. Other provider adapters in the repository are implementation seams, not evidence of interchangeable model capability. No controlled platform-to-platform benchmark, load sweep on live services, or evaluation across environmental families is reported. The communication intervention changes costs, observations, and offered choices together. Six paired habitat seeds and one execution per condition cannot separate service variability from treatment-by-habitat variation. Five pairs happened to run signals-on first; temporal service drift remains a possible confound. Short horizons, engine-assisted navigation, density-dependent births, immigration, and the dispatch throttle constrain interpretation. Numerical-step sensitivity is untested: predation interference is sampled per physics substep without a duration-scaled hazard, so different step partitions can change encounter opportunities. The synthetic timing fixtures suppress predation and cannot bound this pathway. No extra paid condition was added after examining the existing results. The new synthetic validation has its own protocol and does not alter the frozen communication analysis. Source inspection identifies implemented behavior but does not prove that an old deployment used the same revision. Recorded responses support reanalysis of an execution, not bitwise repetition of a changing remote service. Earlier exploratory exports and free-controller runs have weaker source provenance and remain separate. Logs are not a complete world state at every frame. Public archival release, a full dependency/service-license inventory, sponsor-role confirmation, and independent usability evaluation remain unfinished release or research work. 8 Broader impact and disclosure Jevolution can support research on how timing, observations, and interaction mechanisms affect agent evaluation. Its synthetic setting also creates a risk: visually plausible animals and population curves can be mistaken for ecological evidence. All results should be identified as artificial, and ecological applications would require separate calibration and external validation. No live animals or human participants are involved. Any release of prompts, logs, and source must exclude credentials and private information. Bounded request spending and retained failures make the cost of the case study visible; no energy-efficiency claim is made. Author-reported relationships. The original manuscript states that the prototype received Jevathon prizes sponsored by TypeSafe and Cognition, that CodeRabbit provided development resources, and that all three authors interned at CodeRabbit. Sponsor involvement in study design, analysis, and publication decisions is not documented in the available materials. Provenance and assistance. The repository retains the Pixel Agents package identity and an MIT notice crediting Pablo De Lucca.11 1 https://github.com/pixel-agents-hq/pixel-agents Inherited interface infrastructure and assets are not claimed as new contributions. An AI coding assistant supported source inspection, literature verification, experiment tooling, and writing. Authors remain responsible for claims and disclosures. The local artifact contains both this framework framing and an alternative communication-focused draft using the same case-study data; they are not independent studies or two independent empirical contributions. This manuscript is a preprint; no venue acceptance is claimed. References Agarwal et al. [2021] Rishabh Agarwal, Max Schwarzer, Pablo Samuel Castro, Aaron Courville, and Marc G. Bellemare. Deep Reinforcement Learning at the Edge of the Statistical Precipice. In Advances in Neural Information Processing Systems, volume 34, 2021. URL https://proceedings.neurips.c/paper_files/paper/2021/file/f514cec81cb148559cf475e7426eed5e-Paper.pdf. Altera.AL et al. [2024] Altera.AL, Andrew Ahn, Nic Becker, Stephanie Carroll, Nico Christie, Manuel Cortes, Arda Demirci, Melissa Du, Frankie Li, Shuying Luo, Peter Y. Wang, Mathew Willows, Feitong Yang, and Guangyu Robert Yang. Project Sid: Many-agent simulations toward AI civilization. arXiv preprint arXiv:2411.00114, 2024. URL https://arxiv.org/abs/2411.00114. Baker et al. [2020] Bowen Baker, Ingmar Kanitscheider, Todor Markov, Yi Wu, Glenn Powell, Bob McGrew, and Igor Mordatch. Emergent Tool Use From Multi-Agent Autocurricula. In International Conference on Learning Representations (ICLR), 2020. URL https://arxiv.org/pdf/1909.07528. Chopra et al. [2024] Ayush Chopra, Jayakumar Subramanian, Balaji Krishnamurthy, and Ramesh Raskar. flame: A Framework for Learning in Agent-based ModEls. In Proceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 391–399, 2024. URL https://aamas.csc.liv.ac.uk/Proceedings/aamas2024/pdfs/p391.pdf. Efron [1979] B. Efron. Bootstrap Methods: Another Look at the Jackknife. The Annals of Statistics, 7(1):1–26, 1979. doi: 10.1214/aos/1176344552. URL https://blogs.helsinki.fi/bk-club/files/2012/05/Efron_Bootstrap_AS1979.pdf. Foerster et al. [2016] Jakob N. Foerster, Yannis M. Assael, Nando de Freitas, and Shimon Whiteson. Learning to Communicate with Deep Multi-Agent Reinforcement Learning. In Advances in Neural Information Processing Systems, volume 29, 2016. URL https://proceedings.neurips.c/paper_files/paper/2016/file/c7635bfd99248a2cdef8249ef7bfbef4-Paper.pdf. Gao et al. [2024] Chen Gao, Xiaochong Lan, Nian Li, Yuan Yuan, Jingtao Ding, Zhilun Zhou, Fengli Xu, and Yong Li. Large language models empowered agent-based modeling and simulation: a survey and perspectives. Humanities and Social Sciences Communications, 2024. doi: 10.1057/s41599-024-03611-3. URL https://w.nature.com/articles/s41599-024-03611-3. Gause [1934] G. F. Gause. The Struggle for Existence. Williams and Wilkins Company, Baltimore, 1934. URL https://sites.science.oregonstate.edu/~deleenhp/teaching/fall15/MTH427/Gause-The-Struggle-for-Existence.pdf. Grimm et al. [2006] Volker Grimm, Uta Berger, Finn Bastiansen, Sigrunn Eliassen, Vincent Ginot, Jarl Giske, John Goss-Custard, Tamara Grand, Simone K. Heinz, Geir Huse, Andreas Huth, Jane U. Jepsen, Christian Jørgensen, Wolf M. Mooij, Birgit Müller, Guy Pe’er, Cyril Piou, Steven F. Railsback, Andrew M. Robbins, Martha M. Robbins, Eva Rossmanith, Nadja Rüger, Espen Strand, Sami Souissi, Richard A. Stillman, Rune Vabø, Ute Visser, and Donald L. DeAngelis. A standard protocol for describing individual-based and agent-based models. Ecological Modelling, 198(1–2):115–126, 2006. doi: 10.1016/j.ecolmodel.2006.04.023. URL https://faculty.sites.iastate.edu/tesfatsi/archive/tesfatsi/ODDProtocolABM.GrimmEtAl.2006.pdf. Grimm et al. [2010] Volker Grimm, Uta Berger, Donald L. DeAngelis, J. Gary Polhill, Jarl Giske, and Steven F. Railsback. The ODD protocol: A review and first update. Ecological Modelling, 221(23):2760–2768, 2010. doi: 10.1016/j.ecolmodel.2010.08.019. URL https://bio.uib.no/te/papers/Grimm_2010_The_ODD_protocol_.pdf. Grimm et al. [2020] Volker Grimm, Steven F. Railsback, Christian E. Vincenot, Uta Berger, Cara Gallagher, Donald L. DeAngelis, Bruce Edmonds, Jiaqi Ge, Jarl Giske, Jürgen Groeneveld, Alice S. A. Johnston, Alexander Milles, Jacob Nabe-Nielsen, J. Gareth Polhill, Viktoriia Radchuk, Marie-Sophie Rohwäder, Richard A. Stillman, Jan C. Thiele, and Daniel Ayllón. The ODD Protocol for Describing Agent-Based and Other Simulation Models: A Second Update to Improve Clarity, Replication, and Structural Realism. Journal of Artificial Societies and Social Simulation, 23(2):7, 2020. doi: 10.18564/jasss.4259. URL https://w.jasss.org/23/2/7.html. Henderson et al. [2018] Peter Henderson, Riashat Islam, Philip Bachman, Joelle Pineau, Doina Precup, and David Meger. Deep Reinforcement Learning That Matters. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. doi: 10.1609/aaai.v32i1.11694. URL https://ojs.aaai.org/index.php/AAAI/article/view/11694. Huffaker [1958] C. B. Huffaker. Experimental Studies on Predation: Dispersion Factors and Predator-Prey Oscillations. Hilgardia, 27(14):343–383, 1958. doi: 10.3733/hilg.v27n14p343. URL https://my.ucanr.edu/repository/a/?a=152469. Lazaridou et al. [2016] Angeliki Lazaridou, Alexander Peysakhovich, and Marco Baroni. Multi-Agent Cooperation and the Emergence of (Natural) Language. arXiv preprint arXiv:1612.07182, 2016. URL https://arxiv.org/pdf/1612.07182. Leibo et al. [2021] Joel Z. Leibo, Edgar A. Dueñez-Guzman, Alexander Vezhnevets, John P. Agapiou, Peter Sunehag, Raphael Koster, Jayd Matyas, Charlie Beattie, Igor Mordatch, and Thore Graepel. Scalable Evaluation of Multi-Agent Reinforcement Learning with Melting Pot. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 6187–6199, 2021. URL https://proceedings.mlr.press/v139/leibo21a.html. Lotka [1925] Alfred J. Lotka. Elements of Physical Biology. Williams and Wilkins Company, Baltimore, 1925. URL https://archive.org/stream/elementsofphysic017171mbp/elementsofphysic017171mbp_djvu.txt. Lowe et al. [2019] Ryan Lowe, Jakob Foerster, Y-Lan Boureau, Joelle Pineau, and Yann Dauphin. On the Pitfalls of Measuring Emergent Communication. In Proceedings of the 18th International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2019. URL https://ifaamas.org/Proceedings/aamas2019/pdfs/p693.pdf. Masad and Kazil [2015] David Masad and Jacqueline Kazil. Mesa: An Agent-Based Modeling Framework. In Proceedings of the 14th Python in Science Conference (SciPy), pages 51–58, 2015. doi: 10.25080/Majora-7b98e3ed-009. URL https://proceedings.scipy.org/articles/Majora-7b98e3ed-009.pdf. Mordatch and Abbeel [2018] Igor Mordatch and Pieter Abbeel. Emergence of Grounded Compositional Language in Multi-Agent Populations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. doi: 10.1609/aaai.v32i1.11492. URL https://ojs.aaai.org/index.php/AAAI/article/view/11492. Park et al. [2023] Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative Agents: Interactive Simulacra of Human Behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST), 2023. doi: 10.1145/3586183.3606763. URL https://arxiv.org/abs/2304.03442. Piao et al. [2025] Jinghua Piao, Yuwei Yan, Jun Zhang, Nian Li, Junbo Yan, Xiaochong Lan, Zhihong Lu, Zhiheng Zheng, Jing Yi Wang, Di Zhou, Chen Gao, Fengli Xu, Fang Zhang, Ke Rong, Jun Su, and Yong Li. AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society. arXiv preprint arXiv:2502.08691, 2025. URL https://arxiv.org/abs/2502.08691. Ramstedt and Pal [2019] Simon Ramstedt and Christopher Pal. Real-Time Reinforcement Learning. In Advances in Neural Information Processing Systems, volume 32, 2019. URL https://papers.neurips.c/paper_files/paper/2019/hash/54e36c5f5f6a1802925ca009f3ebb68-Abstract.html. Ramstedt et al. [2020] Simon Ramstedt, Yann Bouteiller, Giovanni Beltrame, Christopher Pal, and Jonathan Binas. Reinforcement Learning with Random Delays. arXiv preprint arXiv:2010.02966, 2020. URL https://arxiv.org/pdf/2010.02966. Rosenzweig and MacArthur [1963] M. L. Rosenzweig and R. H. MacArthur. Graphical Representation and Stability Conditions of Predator-Prey Interactions. The American Naturalist, 97(895):209–223, 1963. doi: 10.1086/282272. URL https://w.journals.uchicago.edu/doi/abs/10.1086/282272. Seyfarth et al. [1980] R. M. Seyfarth, D. L. Cheney, and P. Marler. Monkey Responses to Three Different Alarm Calls: Evidence of Predator Classification and Semantic Communication. Science, 210(4471):801–803, 1980. doi: 10.1126/science.7433999. URL https://doi.org/10.1126/science.7433999. Suarez et al. [2019] Joseph Suarez, Yilun Du, Phillip Isola, and Igor Mordatch. Neural MMO: A Massively Multiagent Game Environment for Training and Evaluating Intelligent Agents. arXiv preprint arXiv:1903.00784, 2019. URL https://arxiv.org/abs/1903.00784. Sukhbaatar et al. [2016] Sainbayar Sukhbaatar, Arthur Szlam, and Rob Fergus. Learning Multiagent Communication with Backpropagation. In Advances in Neural Information Processing Systems, volume 29, 2016. URL https://papers.neurips.c/paper_files/paper/2016/hash/55b1927fdafef39c48e5b73b5d61ea60-Abstract.html. Terry et al. [2021] J. K. Terry, Benjamin Black, Nathaniel Grammel, Mario Jayakumar, Ananth Hari, Ryan Sullivan, Luis S. Santos, Clemens Dieffendahl, Caroline Horsch, Rodrigo Perez-Vicente, Niall Williams, Yashas Lokesh, and Praveen Ravi. PettingZoo: Gym for Multi-Agent Reinforcement Learning. In Advances in Neural Information Processing Systems, volume 34, pages 15032–15043, 2021. URL https://papers.neurips.c/paper_files/paper/2021/hash/803f7c4c3f61b71be53a0c803bfb57f-Abstract.html. TypeSafe AI [2026a] TypeSafe AI. Jev 1.13 jaggedness. Official model documentation, 2026a. URL https://docs.typesafe.ai/model-jaggedness/jev-1.13. Provider page last reviewed 17 September 2026; accessed 28 September 2026. TypeSafe AI [2026b] TypeSafe AI. Models. Official API documentation, 2026b. URL https://docs.typesafe.ai/models. Accessed 28 September 2026. TypeSafe AI [2026c] TypeSafe AI. Primitives (Questions). Official API documentation, 2026c. URL https://docs.typesafe.ai/primitives. Accessed 28 September 2026. Volterra [1926] Vito Volterra. Fluctuations in the Abundance of a Species considered Mathematically. Nature, 118:558–560, 1926. doi: 10.1038/118558a0. URL https://w.nature.com/articles/118558a0. Wang et al. [2025] Jianren Wang, Yifan Su, Abhinav Gupta, and Deepak Pathak. Evolutionary Policy Optimization. arXiv preprint arXiv:2503.19037, 2025. URL https://arxiv.org/abs/2503.19037. Wilensky [1999] Uri Wilensky. NetLogo. Center for Connected Learning and Computer-Based Modeling, Northwestern University, 1999. URL https://docs.netlogo.org/copyright. Yu et al. [2024] Yangbin Yu, Qin Zhang, Junyou Li, Qiang Fu, and Deheng Ye. Affordable Generative Agents. arXiv preprint arXiv:2402.02053, 2024. URL https://arxiv.org/abs/2402.02053. Zheng et al. [2018] Lianmin Zheng, Jiacheng Yang, Han Cai, Weinan Zhang, Jun Wang, and Yong Yu. MAgent: A Many-Agent Reinforcement Learning Platform for Artificial Collective Intelligence. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, pages 8222–8223, 2018. doi: 10.1609/aaai.v32i1.11371. URL https://ojs.aaai.org/index.php/AAAI/article/download/11371/11230. Appendix A Recorded-state illustration Figure 4 plots coordinates directly from the retained initial and final records of the first numbered experimental pair. Unlike the historical interface illustration in Figure 2, these states come from the paired study. They illustrate the spatial setting; neither a visually coherent trajectory nor a visible signal establishes that a recipient used a message or that communication improved an outcome. Figure 4: Coordinates from retained world states for pair 01 (habitat seed 80350226): the shared initial state and the two terminal states at 90 simulation seconds. This is a visualization of recorded data, not a screenshot captured during execution. It shows one of the six pairs (n=1n=1 illustrated pair, two runs); the first numbered pair was chosen by index during figure preparation, after data collection, rather than by its effect or appearance. The terrain and initial animal coordinates match across arms. Markers show positions; overlapping markers can hide individuals. Food, energy, signals, and ancestry are omitted. No spatial hypothesis or clustering statistic is tested, so no confidence interval is attached; the aggregate paired estimate and uncertainty are reported in Figure 3. The plot reads initial.json and final.json for main-pair-01-signals_on, and final.json for main-pair-01-signals_off. The renderer verifies the same terrain and starting animal states in both arms. The retained artifact records the input hashes and panel counts in recorded-state-provenance.json; render_recorded_states.py regenerates the figure without calling a provider or advancing the simulation. Appendix B Position relative to existing systems Table 4 lists documented positive properties of neighboring systems. It is a scope comparison, not an exhaustive feature inventory or evidence that an unlisted property is absent. No cross-platform benchmark or ease-of-use experiment was performed. Table 4: Documented system emphases and this paper’s narrower reference task. These are literature/source descriptions, not measured comparative scores. System Documented emphasis Relationship to this implementation NetLogo / Mesa [Wilensky, 1999, Masad and Kazil, 2015] Environments and components for constructing agent-based models. Jevolution supplies one specialized population task and a remote-decision execution/recording path; replacement of general ABM tools is not established. Generative Agents, Project Sid, AgentSociety [Park et al., 2023, Altera.AL et al., 2024, Piao et al., 2025] Agent architectures, real-time social worlds, and simulation/intervention systems. Real-time agent societies and framework-plus-intervention studies are prior art. This case uses a structured decision interface in a mechanistic population task. flame / AgentTorch [Chopra et al., 2024] Differentiable agent-based models for calibration and action learning. Jevolution does not differentiate through its world or optimize its controller in this study. PettingZoo / Melting Pot [Terry et al., 2021, Leibo et al., 2021] Explicit multi-agent interaction API; evaluation across social situations and partners. The present service boundary is source-level and domain-specific; API compatibility and equivalent evaluation breadth are not established. Jevolution Pinned spatial engine, asynchronous requests, population turnover, signals, inherited traits, and joined records. Evidence covers the declared software fixtures and one live-model communication intervention. Practical advantage over the systems above remains an open evaluation question. Appendix C Controller extension and implementation boundaries Concrete controller interface. The frozen RuntimeServices interface supplies identifiers, provider readiness and defaults, an asynchronous choose function, initial-state and event recording, and replay construction. Its decision signature is choose(group, observation, abortSignal, options): Promise<Result>, where group is a ModelGroup, observation is the rabbit/wolf ModelObservation union, and options contains required Boolean relief and predatorPrey flags, optional communication (on/off), and optional requestId. The result carries a legal choice identifier, signal, model identifier, input/output token counts, provider elapsed time, and a nullable provider request identifier. This boundary is defined in core/src/evolution/runtimePorts.ts, lines 18–33; the observation and result types are in core/src/evolution/types.ts, lines 179–260. The following minimal local controller was type-checked against those types. It illustrates interface substitution only; it is not a newly evaluated baseline. The import of RuntimeServices is omitted for space.const localChoice: RuntimeServices[’choose’] = async (_group, observation, _abort, _options) => const rest = observation.choices.find(c => c.id === ’rest’); if (!rest) throw new Error(’No rest option’); return decision: choice: rest.id, signal: ’none’ , model: ’offline-rest-v1’, inputTokens: 0, outputTokens: 0, latencyMs: 0, requestId: null, ; ; Supply this function as choose in the service object passed to new SimulationRuntime(services, auditOptions). An offline host also supplies readiness for its configured groups, deterministic identifiers, and recording/replay services. It should forbid network access explicitly. The engine and observation builder remain unchanged. A more useful controller must handle species-specific observations, choose only supplied actions, and respect the requested communication condition. Virtual time is a testing technique. The frozen service interface has no clock parameter. Existing offline harnesses replace monotonic-clock and timer globals in a separate process and invoke the private runtime tick through a test-only structural cast; pending promise completions must be scheduled explicitly. The earlier free controller harness demonstrates this at scripts/run_free_controls.ts, lines 69–82 and 170–216, in the communication artifact. This avoids engine edits but is a white-box harness, not a stable public virtual-clock API. Global functions are restored afterward. A supported clock port would require a future runtime-interface change. Synthetic timing results therefore describe the specified virtual schedule, not real provider latency. Extensions requiring new implementation. A new species requires extending the species/observation/action types, roster validation, state initialization, observation construction, decision application, lifecycle and runtime dispatch, plus corresponding logging, tests, and any desired interface rendering. The present validator allows only rabbits and wolves (constants.ts, lines 342–395, in the core evolution directory); runtime dispatch distinguishes them by the presence of rabbit genes (runtime.ts, lines 346–359). Adding a provider instead extends the adapter and configuration paths while preserving the decision contract (server/src/evolution/models.ts, lines 115–316). Evolutionary Policy Optimization would require a separate policy representation, evaluation/fitness protocol, policy-variation and selection operators, and persistent generation-level checkpoints and compute accounting. None is supplied by rabbit trait mutation. The current inheritance rule changes five simulator scalars (core/src/evolution/simulation.ts, lines 435–484), while the controller service remains fixed during a run. Such an optimizer would be a new research implementation requiring its own controls and validation. Appendix D Expanded ODD description We use the seven ODD elements [Grimm et al., 2020]. The experimental implementation is based on commit 9285979e95d67d8f11fa4f6d9b603b156c84380, with the frozen patch identified in Appendix F. It corrects reproduction instructions, enforces communication-off at the engine, adds logging and spending controls, and bounds dispatch frequency. Historical exports do not establish that they were generated by precisely this revision and are not pooled with the new study. D.1 Purpose and patterns The model is an artificial decision-making task for examining how local information and asynchronous execution affect simulated population outcomes. The completed remote-controller evaluation concerns a short-horizon communication intervention; the framework also receives separate synthetic software validation. It does not target any measured rabbit or wolf population, fit ecological parameters, or predict conservation outcomes. Visual clustering, repeated population fluctuations, and trait shifts are not validation criteria without additional controls. D.2 Entities, state variables, and scales The reference habitat is a 64×6464× 64 grid with terrain and food resources. Animal state includes identity, position, energy, age, lineage, current activity, and decision scheduling. Rabbits also carry bounded memory and five scalar traits: speed, vigilance, thrift, fertility, and sociability. Signals store their source, location, type, delivery time, and expiry. Terrain is fixed during an unedited run; food and animal states evolve. Simulation time and wall time are retained separately in the logs. D.3 Process overview and scheduling The engine advances resources, movements, metabolic state, predation, reproduction, and immigration while model requests are outstanding. A request packages the animal’s local observation and offered actions; a returned choice is normalized and may be rejected or no longer applicable. In the reference runtime, the 2,5002,500-ms acceptance threshold is measured in wall time, while the default 100100-ms post-request cooldown is measured in simulation time and permits engine-triggered eligibility resets. The reference concurrency limit of six is per lineage, not per animal or globally. The transport timeout is separate from the acceptance threshold. Exact process order, fallback behavior, invalid-choice handling, and any experimental dispatch throttle are part of the configuration, not incidental implementation details. D.4 Design concepts Agents have partial observations, but engine functions may consult the full map when generating legal routes. This distinction prevents attributing all feasible navigation to the model. Signals use a fixed vocabulary: danger, food, follow, and silence in the predator–prey scenario; the help mechanism is disabled and wolves do not signal. The provider selects structured choices from the offered actions and signals, rather than composing an unrestricted message. A received message does not mechanically force a response. Visually supported following can occur without a message and remains available in the off condition. The evaluated controller is TypeSafe’s Jev jev-1.13.0, which accepts text or structured text state and returns typed decisions [TypeSafe AI, 2026b]. The provider describes this version as a decision model without text-generation training [TypeSafe AI, 2026a]. Action and signal use separate Choice questions over one shared observation. The adapter uses each returned choice; returned confidence and probability fields are logged but do not gate, weight, or resample decisions. Appendix H gives the request schema and code anchors. The study did not independently inspect model architecture, training data, or weights, and its result concerns this served decision interface. The controller is not trained within a run. A fixed requested model identifier records the intended remote policy; it cannot independently prove that a vendor’s underlying weights remained immutable. Rabbit offspring inherit mutated traits, while immigrants receive new traits. Consequently, the living population’s trait distribution can change through reproduction, mutation, mortality, immigration, and stochastic variation. A shift alone does not establish adaptive selection. D.5 Initialization The predator–prey preset creates 70 rabbits and 8 wolves. Pairing uses identical treatment-invariant initial biological state: habitat, resources, animal IDs, positions and traits, and initial engine random-number state. A separate full-condition hash records the intervention. Equal seeds alone do not guarantee identical later random events after action choices and request completion order diverge. Terrain, initial animal state, and hashes are saved rather than reconstructed from an unversioned preset name. D.6 Input data The experiment supplies remote model responses rather than observations of real animals. Append-only logs retain requests, returned model identifiers, token usage, errors, cancellations, timestamps, and normalized choices, with shared request IDs. Population counts are logged after each engine substep, which is at most 50 simulation ms; complete initial/final states and approximately one-second world snapshots are also retained. This is not a full-state replay at every frame. Substep endpoint differences can miss an animal born and killed within that substep, although cumulative demographic counters retain those events. API credentials are excluded from the artifact. No empirical ecological calibration data enter this prototype. D.7 Submodels For sociability s, the reference signal radius and immediate energy cost are r(s)=3+5s,c(s)=0.6(1+s).r(s)=3+5s, c(s)=0.6(1+s). (3) Delivery is scheduled 0.20.2 simulation seconds after emission; expiry is six simulation seconds after emission. The global history retains at most 200 signals; oldest-entry eviction can end availability earlier than the nominal expiry. Thus, absent history eviction, the nominal available interval is 5.85.8 seconds, and intermittent observations can shorten actual exposure. Range, terrain visibility, and sender identity further restrict reception. The signal location is its emission location. In this preset, reproduction is automatic and single-parent for both species. For a rabbit trait g, the reference inheritance rule is gchild=clip[0.05,0.95](gparent+ϵ),ϵ∼[−0.065,0.065),g_child=clip_[0.05,0.95](g_parent+ε), ε [-0.065,0.065), (4) where the implementation uses a finite pseudorandom generator. This describes a simulator parameter mutation, not a policy-gradient update. Energy, age, capacity, placement, and dispersal rules constrain births. Rabbit requests describe automatic reproduction, and the experimental patch removes the earlier contradictory mutual-mating option. Food follows capped linear regrowth with terrain-dependent rules; it is not a logistic population equation. The predator–prey preset uses reference food capacity 8 and normal regrowth coefficient 0.1. No drought is triggered in the study. Immigration attempts add two rabbits and one wolf after at least 20 simulation seconds since the prior eligible check, subject to placement. These arrivals are separate from births and may occur after a species reaches zero. Between resets and external edits, population accounting requires N(t)=N(0)+B(t)+I(t)−D(t),N(t)=N(0)+B(t)+I(t)-D(t), (5) with births B, immigration I, and deaths D. A zero count is therefore not necessarily permanent extinction. Appendix I specifies effective recruitment and predation parameters and the ordered physics substeps, including density-dependent and stationary-only rabbit breeding. These are artificial task settings, not calibrated biological rates. Appendix E Synthetic validation design and complete-attempt reporting These fixtures are separate from the remotely served communication study. They make no network calls and load no credentials. Runtime fixtures use a deterministic rest/none controller labeled synthetic-fixture, with zero provider tokens; signal/reproduction fixtures invoke their named mechanisms directly. Source hashes and a protocol were saved before the execution sequence. The protocol, runner versions, full records, and checks are retained under validation/. Clock and runtime matrix. The runner crosses seeds 42 and 314159, response delays 0, 100, 500, 1,000, 1,001, and 1,500 ms, and outstanding capacities 1 and 3 per lineage. The fixture uses realtime scheduling, a 1,000-ms acceptance deadline, 100-ms post-request interval, no additional global throttle, and a 4.137-s horizon. One simulation second corresponds to 1,000 virtual wall ms. A 10-ms grid is augmented by response/timer events; physics ticks precede response resolution at the same timestamp. Virtual time uses the white-box harness described in Appendix C, not a public clock API. The predator–prey preset is used as a container and then deliberately edited: empty grass terrain, three rabbits and three wolves placed far apart, initial energy/water 90, age zero, no initial paths, and breeding/immigration suppressed. These modified fixtures are not ecological replications. An applied observation’s age is checked against the injected delay. Count identities require requests = applied + late + invalid + errors + cancelled; caps, complete joins, terminal settlement, and exact rerun hashes are checked. A pause at virtual wall time 300 ms and removal of a pending rabbit at that time separately exercise cancellation. Pause occurs before the next physics tick, so its recorded simulation horizon is 0.29 s; it is an intentional stop, not a failed 4.137-s run. Signals. At sociability 0.05, 0.50, and 0.95, a stationary sender with ample energy emits follow, or attempts it with signaling disabled. Receivers are positioned at radius minus 0.001, exactly at radius, and radius plus 0.001, and observed at 0.199, 0.200, 5.999, and 6.000 simulation seconds. All 72 on/off observation cells are retained. Checks cover the implemented cost and radius, inclusive range boundary, delivery eligibility at 0.2 s, and expiry at 6 s. Off-mode emission is rejected without changing energy or signal counters. Terrain occlusion and model listening are not tested. Inheritance and arrivals. Seeds 1–16 are crossed with three parental gene profiles (both 0.05; 0.20 and 0.80; both 0.95) and arena versus predator–prey scenarios. The two mature, well-fed, same-lineage parents have generations 2 and 5. Arena uses reciprocal mate actions; predator–prey retains automatic single-parent reproduction. Wolves, immigration, and unrelated spatial constraints are suppressed before one 50-ms physics step. All 192 executions remain in the inventory. Across the 96 first executions, there are 141 offspring records and 0 executions without a birth; exact repeats are excluded from these two counts. These are fixture counts, not estimated reproductive rates. For each child, its recorded parent IDs determine the appropriate parent mean; its generation must be one greater than the largest recorded parent generation. Each gene must be in [0.05,0.95] and within 0.065 of that parent mean before clipping. Repeated initial and final states are hash-compared. Checking bounds and repeatability does not test the full mutation distribution or establish fitness differences. A separate repeated empty-population fixture for the same 16 seeds crosses the immigration timer and checks two rabbits plus one wolf, empty parent lists, generation zero, and counters. Generation zero alone does not distinguish a founder from an immigrant; timing and arrival records provide that context. Attempt history and scope. An initial launcher encountered the sandbox’s IPC restriction before the runner executed; a Node loader entrypoint avoided that limitation. The first executed harness then stopped at the terminal boundary because floating-point simulation time was just below the horizon while virtual wall time already equaled it. Partial records and the original runner are retained. The correction lets virtual time advance so existing horizon clipping completes; source, factors, and criteria are unchanged. The completed attempt reports 2,724 passing checks and no failing check. A clean full-process rerun matched every runtime summary and all saved signal, inheritance, immigration, and invariant arrays. Separate raw-record arithmetic checks also passed; neither is a new biological sample. No paid API call was made. Appendix F Frozen protocol and artifact identifiers The main-study record was frozen locally at 2026-09-28T23:30:56.062648+00:00. Its SHA-256 is d023029c8288e9c7571667ac9b2e542dc6fe49c5b8189cb1c693c14bce8e791 The source base is commit 9285979e95d67d8f11fa4f6d9b603b156c84380. The applied implementation patch has SHA-256 98135f920a582b2b6c3117340cdcceea6cedaffc5b7c9424451fef727c956bfb These identities distinguish the experimental implementation from the historical illustrative exports and from the unused newer branch. The local bundle contains the patch, dependency lockfile, effective rules, source/script hashes, and study-design manifest. No independently archived dataset DOI is claimed. Table 5: Frozen main-study seed blocks and order. Six pairs were planned, with one attempted run per arm and no outcome-selected replacements. Actual completion is reported in the run ledger. Pair Habitat seed First arm 1 80350226 Signals on 2 1766849094 Signals on 3 1976771215 Signals off 4 865173864 Signals on 5 1328634472 Signals on 6 415864476 Signals on Seeds were sampled without replacement from integers 1 through 2147483647 with Python random.Random(2026092801). First-arm bits used a separate generator initialized with 2026092802. The realized five-to-one order imbalance was not rerandomized. The primary analysis generator uses seed 20260928. These identifiers specify the intended randomization rather than guarantee identical asynchronous remote-model outputs. Stopping and provenance. The hard horizon is 90 simulation seconds and the run request ceiling is 902. The global 100-wall-ms dispatch spacing bounds submissions to ten per wall second; the ceiling provides operational headroom and is not an invitation to exceed the horizon. The runtime continues through temporary empty-population states. Requests pending at the horizon are cancelled and their unknown usage is conservatively accounted for. The runtime does not automatically retry a failed transport request. Every later dispatch is separately logged and reserved. Paired inference requires complete horizons, matching treatment-invariant initial-state hashes, and matching source/configuration/protocol provenance. The integration rule uses every recorded engine substep and never substitutes an observed shorter horizon. The complete-pair estimate is accompanied by exclusions and stop reasons, not silently selected successful trajectories. Recorded clarification. After freezing, source review corrected the protocol’s description of the 100-simulation-ms animal delay: it is a post-request cooldown with engine-triggered eligibility resets, including hiding transitions and wolf births. The strict global 100-wall-ms dispatch spacing is unchanged. This was a documentation correction; no executable, intervention, seed, endpoint, or analysis changed, and no interim treatment effects were used. The frozen protocol and its hash remain intact, with the correction recorded separately. Appendix G Complete attempt inventory Every recorded attempt is included below. P denotes paid main pairs, H the handwritten reference, and +/−+/- signals on/off. Pilot and engineering labels follow the analysis inventory order. The machine-readable analysis retains exact run identifiers, input hashes, and stop reasons. The handwritten controller uses immediate decisions and ignores received messages; its outcomes are descriptive, not a latency-matched model comparison. Earlier pilot/free-control patch hashes are retained, but their superseded patch bytes were not independently archived; the final frozen source archive establishes main-study provenance. Table 6: All attempted runs. L/I/E/C denotes late, invalid, error, and cancelled requests. Usage and per-run conservative bounds are derived from the paid-request ledger; free/engineering runs make no paid calls. These are not per-run invoices. Times are observed simulation horizons. Run Time (s) Req. Applied L/I/E/C Usage ($) Bound ($) Status E1+ 2.0 20 20 0/0/0/0 0.0000 0.0000 complete E2- 2.0 20 20 0/0/0/0 0.0000 0.0000 complete H01- 90.0 900 900 0/0/0/0 0.0000 0.0000 complete H01+ 90.0 900 900 0/0/0/0 0.0000 0.0000 complete H02- 90.0 900 900 0/0/0/0 0.0000 0.0000 complete H02+ 90.0 900 900 0/0/0/0 0.0000 0.0000 complete H03- 90.0 900 900 0/0/0/0 0.0000 0.0000 complete H03+ 90.0 900 900 0/0/0/0 0.0000 0.0000 complete H04- 90.0 900 900 0/0/0/0 0.0000 0.0000 complete H04+ 90.0 900 900 0/0/0/0 0.0000 0.0000 complete H05- 90.0 900 900 0/0/0/0 0.0000 0.0000 complete H05+ 90.0 900 900 0/0/0/0 0.0000 0.0000 complete H06- 90.0 900 900 0/0/0/0 0.0000 0.0000 complete H06+ 90.0 900 900 0/0/0/0 0.0000 0.0000 complete P01- 90.0 879 876 0/1/0/2 0.0963 0.1018 complete P01+ 90.0 878 874 0/3/0/1 0.1019 0.1046 complete P02- 90.0 856 844 0/5/4/3 0.0928 0.1121 complete P02+ 90.0 879 872 0/5/0/2 0.1028 0.1083 complete P03- 90.0 881 873 0/4/0/4 0.0967 0.1077 complete P03+ 90.0 877 874 0/3/0/0 0.0971 0.0971 complete P04- 90.0 878 876 0/1/0/1 0.0941 0.0968 complete P04+ 90.0 875 874 0/0/0/1 0.0947 0.0975 complete P05- 90.0 878 876 0/1/0/1 0.0935 0.0962 complete P05+ 90.0 878 876 0/1/0/1 0.0975 0.1003 complete P06- 90.0 878 876 0/2/0/0 0.0979 0.0979 complete P06+ 90.0 877 871 0/5/0/1 0.0996 0.1023 complete Pilot1+ 30.0 290 288 0/0/0/2 0.0293 0.0348 complete Pilot2+ 30.0 294 290 0/3/0/1 0.0319 0.0346 complete Table 7: Secondary descriptive outcomes for all 12 main attempts. Wolf AUC is normalized by initial wolf abundance; rabbit founder survival is the fraction of initial rabbit IDs remaining at the fixed horizon. No secondary causal mechanism is identified and no extra hypothesis test is performed. Run Wolf AUC Founder surv. R births R deaths R arrivals Signals P01- 3.3617 0.400 141 103 8 0 P01+ 3.1830 0.500 151 106 8 191 P02- 3.8189 0.386 149 121 8 0 P02+ 3.3332 0.457 143 104 8 169 P03- 3.1411 0.429 147 104 8 0 P03+ 3.0703 0.457 135 104 8 222 P04- 3.6287 0.414 140 120 8 0 P04+ 3.7298 0.386 135 113 8 220 P05- 3.4950 0.514 160 104 8 0 P05+ 3.5623 0.429 144 108 8 211 P06- 2.5380 0.614 129 70 8 0 P06+ 2.7092 0.600 131 72 8 155 Across all main runs, cumulative engine counters record 1229 rabbit deaths: 1229 attributed to predation and 0 to starvation. These are simulator event categories, not independently observed biological causes. Table 8: Handwritten reference outcomes for 6 seed blocks. These are descriptive normalized rabbit-AUC values with immediate, zero-provider-latency decisions. The policy ignores received signals; its on/off differences can reflect emission cost. These runs are excluded from the main estimate and are not a matched comparison of model capability. Pair On AUC Off AUC On −- off pair-01 1.6974 1.7245 -0.0271 pair-02 1.6762 1.6677 0.0085 pair-03 1.6028 1.5810 0.0218 pair-04 1.6469 1.6192 0.0277 pair-05 1.6734 1.6774 -0.0039 pair-06 1.7273 1.7530 -0.0258 Figure 5: All 6 complete main-study seed pairs (12 runs), grouped by the frozen habitat seed. Each panel compares the two recorded arms for one initial state; provider responses and subsequent stochastic paths may differ. These descriptive traces do not establish ecological cycles or message-use mechanisms. Appendix H Evaluated controller interface The completed study uses TypeSafe’s jev-1.13.0 through its /v1/systemone endpoint. The provider documents text inputs, typed decisions, and independently evaluated questions sharing one state [TypeSafe AI, 2026b, TypeSafe AI, 2026c]. Its version-specific documentation states that Jev is not trained for text generation [TypeSafe AI, 2026a]. These are provider descriptions of the served interface. The artifact does not supply model weights, a training corpus, or an independently verified architecture. The experiment therefore characterizes the recorded controller and runtime; it does not evaluate a generative chat-agent architecture. Request structure. The following schematic shows the two questions for a signals-on rabbit. Placeholder state fields, action identifiers, instructions, and abbreviated signal descriptions illustrate the structure; this is not an archived request or the exact experimental prompt. Full requests, including actual legal options and instructions, are retained in each run’s requests.jsonl. "model": "jev-1.13.0", "state": "...": "local observation fields", "questions": "action": "type": "choice", "instructions": "<scenario and condition instructions>", "criteria": "<action ID>": "<legal action description>" , "signal": "type": "choice", "instructions": "<same instructions; select a signal>", "criteria": "none": "<silence criterion>", "danger": "<warning criterion>", "food": "<food criterion>", "follow": "<following criterion>" Both questions see the same observation. Under the documented interface, neither question sees the other’s answer. In particular, the signal choice is not conditioned on the action selected in that request. The code combines the two returned choices into one decision for the engine; no second inference call coordinates them. Wolves and communication-off rabbits receive only the none signal option. This batching preserves one request/application opportunity, but does not establish coordinated optimization over action–signal pairs. Response handling and source anchors. The frozen server/src/evolution/models.ts, lines 165–179, constructs the two questions; lines 278–298 extract answers.action.choice and answers.signal.choice. Lines 99–113 validate these values against the offered action and signal sets. The returned confidence and per-option probabilities are retained in provider-response logs, but the adapter does not threshold, weight, or sample from them. The runtime then checks cancellation, lateness, and current applicability before applying the combined decision (core/src/evolution/runtime.ts, lines 421–445). These checks concern request timing and engine validity; they are not confidence gates. This description clarifies the frozen implementation and does not modify the policy used to collect the study. Appendix I Population submodels and effective task parameters This supplement expands the ODD summary for the frozen main-study implementation. Values in Table 9 are arbitrary simulator task settings, not estimates calibrated to animal populations; no biological or optimality rationale is inferred from them. All 12 main-run metadata records contain the same effective rules. Time below is simulation time, distance is in grid tiles, and energy, water, and food use simulator units. This is a readable specification of the mechanisms most directly affecting the abundance endpoint, not a replacement for the complete frozen code, terrain-generation procedure, or legal-action specification. The companion MECHANISM-APPENDIX-EVIDENCE.md maps every rule below to source lines and retained configuration. Table 9: Effective parameters governing movement and demography. Trait symbols gs,gv,gt,fg_s,g_v,g_t,f denote speed, vigilance, thrift, and fertility. Population caps constrain births, not immigration. Mechanism Main-study setting Rabbit recruitment Maturity 8 s; minimum energy and water 30 each; global crowding scale 160 rabbits; hard birth cap 250; local crowding disabled. Rabbit birth cost Parent loses 10+12f10+12f energy; birth or crowding rejection resets cooldown to 16(1.3−0.6f)16(1.3-0.6f) s. Newborn energy/water: 60/80. Wolf recruitment Maturity 5 s; energy threshold 80; parent cost 30; birth cooldown 10 s; initial founder birth delay 20 s; birth cap 150. Newborn energy: 33. Wolf energy Metabolism 1.2 energy/s; a kill adds 40, capped at 100; movement energy charge is zero. Movement Rabbit speed 1.05+0.65gs−0.16gt1.05+0.65g_s-0.16g_t tiles/s; wolf speed 1.6 tiles/s. Forest multiplies speed by 0.77; a wolf’s eating cooldown adds a factor 0.35. Detection and capture Rabbit sight radius 4+⌊6gv⌋4+ 6g_v ; wolf sight radius 9, or 2 for a rabbit hiding in forest. Capture distance <0.8<0.8; eating cooldown 10 s; interference coefficient 1. Newborn dispersal Up to 12 route attempts; proposed radial distance in [6,12)[6,12) before boundary clipping and tile centering. Actual displacement and route length need not lie in that interval. Resources Reference food capacity 8; nondrought regrowth coefficient 0.1; food conversion 5.5 energy/unit; rabbit thirst 0.1 water/s. Lifespan and arrivals Both lifespan thresholds 10910^9 s. Every eligible 20-s immigration check attempts two rabbits and one wolf, subject to placement. Ordered engine substep. The runtime splits elapsed simulation time into substeps of at most 0.05 s, clipping the last substep at the study horizon. Returned model decisions are handled asynchronously; the following is the order within each synchronous physics substep, not an assumption that all animals decide together. 1. Advance the simulation clock, regrow food, and expire signals (or clear them when signaling is off). On a food-bearing tile of capacity K, nondrought regrowth adds 0.1(K/8)dt0.1(K/8)\,dt, capped at K. 2. Copy the current rabbit list and rotate its order by a pseudorandom offset. For each rabbit in that order, advance age and cooldown; debit basal energy and water; update following and hiding; move and charge movement energy. If its path is empty, process stationary activities, including foraging, drinking, and automatic breeding. Then remove that rabbit if energy or water is exhausted, or age exceeds its threshold. 3. Process a copy of the current wolf list in its stored order. Advance age and debit metabolism, update cooldowns and pursuit, move, check for a capture, credit a meal if caught, and then remove wolves with exhausted energy or excessive age. 4. Attempt wolf reproduction, then immigration; finish any drought transition and update the bounded display history. The experiment records population counts after the completed substep. Rabbit newborns are appended during step 2 and are not themselves updated in that copied rabbit loop; they can nevertheless be encountered by wolves in step 3. Wolf newborns enter after the wolf loop. Consequently, births, resource use, and predation are sequential rather than simultaneous events. Rabbit recruitment and energy. A rabbit attempts automatic single-parent reproduction only when its path is empty, age is at least 8 s, cooldown is nonpositive, energy and water are each at least 30, and no wolf is within four tiles (this last test uses distance, not visibility). Let R be the current rabbit count at the attempt. The hard cap first rejects attempts with R≥250R≥ 250. Otherwise a finite-generator draw u∈[0,1)u∈[0,1) rejects the attempt when u<R/160u<R/160; because the draw lies in [0,1)[0,1), R≥160R≥ 160 deterministically rejects a birth attempt even below the hard cap of 250. Such a rejection resets the parent’s fertility-dependent cooldown without charging birth energy. This is an eligibility-and-cooldown rule, not a continuous logistic population equation. A surviving attempt requires an adjacent walkable location unoccupied by another rabbit; failed placement does not apply the birth cost or cooldown. A successful birth applies the cost and cooldown in the table and clears the parent’s current movement/action to rest. Rabbit basal energy loss per second is 0.1+0.06gv+0.05gs−0.035gt0.1+0.06g_v+0.05g_s-0.035g_t, multiplied by 0.7 while resting. Movement costs 0.1+0.14gs0.1+0.14g_s energy per tile. Stationary foraging consumes up to 1.4(1−0.3gt)1.4(1-0.3g_t) food units/s, limited by local food and the energy ceiling of 100; stationary drinking beside water restores up to 16 water units/s, capped at 100. Thus an action can affect abundance through reproduction eligibility and energy as well as predation. This is a possible implemented pathway, not a mechanism identified by the treatment estimate. Refuges, pursuit, and wolf recruitment. Shelter tiles exclude rabbit detection by wolves and exclude capture. Forest hiding reduces detection range but is not absolute protection from an encounter. Visibility additionally tests terrain obstruction. Model-controlled wolves replan an active hunt at intervals of at least 0.65 s while the selected prey remains visible. A wolf can also catch another exposed rabbit it reaches. For a candidate capture with n other wolves within four tiles, the interference check passes when n=0n=0 or a pseudorandom draw is below 1/(1+n)1/(1+n). This check is made within substeps; it is not a fitted continuous-time predation rate. Capture additionally requires a nonpositive eating cooldown, then resets it to 10 s. Wolf births are automatic and single-parent, requiring the age, energy, cooldown, and cap conditions in the table plus adjacent non-shelter placement unoccupied by any animal. Unlike rabbit births, they do not require the parent to stop moving. Founder wolves begin mature, with energy 50, an eating cooldown of 8 s, and a 20-s birth cooldown. Rabbit founders begin mature, with energy/water 75/90 and an 8-s birth cooldown. Newborns of both species receive dispersal routes if an attempted destination is legal and reachable. Dispersal sets the next decision time to current time plus straight-line destination distance divided by 1.2; this is an eligibility estimate, not a guarantee that the animal has arrived. Open-population boundary. Immigrants are placed beside conspecifics when possible, with an edge-tile fallback. They bypass the birth caps. Arriving rabbits receive fresh trait draws, mature age, and founder resource values; arriving wolves receive founder state except that their birth cooldown is reset to zero. The high lifespan thresholds disable old-age removal over this 90-s horizon. These choices make abundance an outcome of an explicitly open artificial population; they do not establish long-run coexistence, fitness, or a reduction to a classical predator–prey equation. Appendix J Post-hoc timing and signal-exposure diagnostics These descriptive diagnostics were added after collection in response to reporting review. They use all 12 complete main runs and change no frozen endpoint, inclusion rule, or uncertainty estimate. Rabbit observations contained at least one signal in 828/2634 on-condition requests (31.44%) and 0/2601 off-condition requests. Exposure means a signal was present in the supplied snapshot; it does not establish attention, listening, behavioral use, or unique delivery events. Queue delay runs from engine eligibility to dispatch in simulation seconds. Observation age instead runs from the captured observation to application, also in simulation seconds and only for applied decisions. Recorded response time is provider-adapter elapsed wall time for available returns, including transport and local handling; it is not isolated model compute. Missing latency is not imputed as zero. Medians and 95th percentiles use linear interpolation at (N−1)p(N-1)p. Decision-opportunity rates divide requested observations by total living-species animal-minutes, integrated with the frozen left-hold rule and including infants and other animals not yet eligible to act. Thus these are population exposure-normalized rates, not each eligible adult’s decision frequency. Decisions and time samples are not independent replicates; no confidence interval, effect test, or causal latency comparison is added. Table 10: Signal exposure and population-normalized opportunities. +/−+/- denotes signals on/off. Type columns count rabbit observations containing each kind and are nonexclusive; help is absent. The final columns are rabbit/wolf requests per living animal-minute, not per wall minute or per eligible adult. Run Obs. Any % Danger Food Follow R req/min W req/min P01- 440 0 0.0 0 0 0 2.65 10.88 P01+ 439 148 33.7 105 46 0 2.44 11.49 P02- 403 0 0.0 0 0 0 2.46 9.88 P02+ 440 121 27.5 97 25 0 2.56 10.98 P03- 441 0 0.0 0 0 0 2.56 11.67 P03+ 439 162 36.9 146 18 2 2.70 11.89 P04- 439 0 0.0 0 0 0 2.75 10.08 P04+ 438 133 30.4 116 15 6 2.78 9.76 P05- 439 0 0.0 0 0 0 2.56 10.47 P05+ 439 149 33.9 116 41 0 2.76 10.27 P06- 439 0 0.0 0 0 0 2.50 14.41 P06+ 439 115 26.2 105 11 0 2.54 13.47 Table 11: Per-run/species timing: R denotes rabbits and W wolves. Q is eligibility-to-dispatch queue delay; W in the time header is provider-adapter elapsed wall time. Time entries give median/95th percentile. Queue values are available for every listed observation. Return n/N explicitly excludes missing latency from quantiles; Age n/applied gives the matched-age denominator among applied requests. These are descriptive conditional distributions, not independent experimental samples. Run/species Obs. Q (sim s) Return n/N W (wall ms) Age n/applied Age (sim s) P01- R 440 18.24/19.49 438/440 112.5/172.5 438/438 0.102/0.153 P01- W 439 3.46/7.80 439/439 107.5/154.9 438/438 0.101/0.152 P01+ R 439 19.03/22.98 439/439 116.0/170.9 439/439 0.102/0.153 P01+ W 439 3.11/7.63 438/439 107.3/152.1 435/435 0.101/0.152 P02- R 403 17.36/24.82 398/403 112.9/161.8 398/398 0.102/0.153 P02- W 453 3.64/9.44 451/453 108.6/156.8 446/446 0.101/0.152 P02+ R 440 18.38/20.49 438/440 114.8/162.0 438/438 0.102/0.153 P02+ W 439 3.32/7.75 439/439 109.1/155.8 434/434 0.101/0.152 P03- R 441 18.11/21.39 437/441 116.4/164.7 437/437 0.102/0.153 P03- W 440 3.47/7.14 440/440 112.5/165.0 436/436 0.101/0.153 P03+ R 439 17.39/20.01 439/439 116.9/165.8 438/438 0.102/0.153 P03+ W 438 3.51/7.50 438/438 112.4/161.5 436/436 0.102/0.153 P04- R 439 17.40/18.85 439/439 115.7/157.1 439/439 0.102/0.152 P04- W 439 3.91/9.10 438/439 112.8/160.7 437/437 0.101/0.152 P04+ R 438 17.14/19.21 437/438 117.6/161.0 437/437 0.102/0.153 P04+ W 437 3.78/8.87 437/437 111.5/160.5 437/437 0.101/0.152 P05- R 439 18.02/20.13 439/439 118.3/183.9 439/439 0.102/0.153 P05- W 439 4.02/8.73 438/439 116.8/169.5 437/437 0.102/0.153 P05+ R 439 17.56/19.02 439/439 119.5/169.0 439/439 0.102/0.153 P05+ W 439 3.50/9.09 438/439 115.4/160.9 437/437 0.101/0.152 P06- R 439 17.79/22.86 439/439 121.2/167.1 439/439 0.102/0.153 P06- W 439 3.20/5.52 439/439 118.7/158.6 437/437 0.102/0.152 P06+ R 439 18.39/21.71 438/439 117.7/159.5 438/438 0.102/0.152 P06+ W 438 3.26/6.33 438/438 115.5/152.1 433/433 0.101/0.105 Appendix K Artifact boundaries and reuse This manuscript is a framework-oriented alternative to a communication-focused draft prepared from the same Jevolution audit. The communication logs, frozen design, primary estimator, and reported effect are shared evidence. They are not new replications, a second independent study, or grounds for counting the same empirical contribution twice. The new evidence consists of separately declared free synthetic engineering validation and the source-backed architectural specification. The authors retain a local research package containing the original communication artifact, the separate framework source, the new validation protocol/scripts/results, claim and citation ledgers, and a comparison of the two framings. This package is distinct from the manuscript’s typesetting source and is not publicly archived at this revision. The original communication manuscript and frozen inputs remain unchanged. One command regenerates framework figures and compiled manuscripts from these retained inputs; a separately documented free command reruns the engineering suite. Neither invokes remote inference. These are local reproduction checks, not a claim that readers currently have access to the complete package. The source-backed description covers the pinned experimental implementation. Historical UI exports and the live website are not substituted for that provenance. Bounded historical journals cannot establish complete decision frequencies. The communication run inventory includes all earlier attempts; the synthetic fixtures have a separate attempt inventory and are never pooled into its sample size or uncertainty calculation. Project links and experimental version. The project website is https://jevolution.world, and the public source repository is https://github.com/utkarshg20-org/jevolution. The recorded base revision is 9285979e95d67d8f11fa4f6d9b603b156c84380. A separate retained patch and frozen source archive identify the evaluated implementation (Appendix F). These links identify the project and base source; they do not substitute for the complete experimental supplement or certify the revision running on the website. Documentation accompanying the new assets. The retained package’s ASSET-CARD.md documents the code, synthetic records, shared communication logs, schemas, provenance, intended uses, limitations, and license status. Its README.md gives the exact analysis/build and optional fixture commands. Raw records retain failures and distinguish simulation time from wall time; runtime fixtures are software checks, not independent ecological samples. No new model weights or human-subject dataset is introduced. The arXiv typesetting package contains the manuscript and its figures; it does not distribute the frozen engine, raw service responses, or full experiment dataset. A public repository link alone therefore does not provide independent reproduction of this exact study. Asset provenance and license scope. The frozen source retains the Pixel Agents MIT notice, copyright 2026 Pablo De Lucca. Project records identify the animal and terrain atlases as AI-generated artwork created on 24 September 2026; inherited interface infrastructure and generated artwork are distinct assets. The interface uses FS Pixel Sans Unicode by NZWStudios2024 under SIL Open Font License 1.1.22 2 Font and license: https://fontstruct.com/fontstructions/show/2606508/fs-pixel-sans-unicode-regular. The historical screenshot in Figure 2 illustrates the authors’ original manuscript; it does not establish the deployed code or artwork revision. The architecture, statistical plots, and recorded-state map are generated by the retained scripts. Jev is a hosted service subject to provider terms. The currently posted TypeSafe agreement discusses output ownership and also limits specified downstream uses; the account-specific terms governing the experiments have not been verified.33 3 Current terms inspected 1 October 2026: https://typesafe.ai/legal/mca. No unrestricted license for the raw-response dataset is asserted. The MIT source notice, interface-font license, and manuscript-distribution license must not be interpreted as a blanket license for every dependency, artwork, or service output. A complete redistribution review remains separate from reporting aggregate research results. The full research package remains local. A future release should attach a persistent archive identifier and finalize asset/service and data licensing. Sponsor roles and a future venue’s requirements must also be confirmed. An anonymous PDF is not an anonymized raw-data bundle. No public dataset DOI, independent reuse, or venue acceptance is asserted. Appendix L Execution resources and accounting The recorded host was an Apple M5 Pro with 15 logical CPUs, 24 GiB installed memory (25,769,803,776 bytes), macOS (Darwin 25.6.0), and arm64 architecture. Software versions were Node 26.8.1, TypeScript 5.9.3, and tsx 4.23.15. Synthetic manifests independently confirm Node/platform/architecture, without a separate hardware fingerprint. These are host resources, not peak process requirements. Local Node runs the simulator; TypeSafe’s remote jev-1.13.0 API supplies paid decisions. No policy training occurs. Paid runs execute serially, with at most six pending requests per lineage and at least 100 wall-ms between shared dispatches. Table 12 reports real UTC end-minus-start intervals for all 28 original attempts: 1,153.393 s (19.223 min) total, comprising 1,081.125 s for 12 main runs, 60.246 s for two pilots, 7.900 s for 12 offline handwritten controls, and 4.122 s for two engineering checks. These are elapsed intervals, not CPU-core-seconds, exclusive kernel time, or whole-project duration. The controls’ virtual clock advances 90 simulation seconds faster than real time. The paid runs made 11,098 API requests; usage-based cost is $1.226087394 and conservative accounting including unknown usage is $1.292147682, not an independently verified invoice. Table 12: Recorded elapsed seconds for every original attempt. P: paid main; H: offline handwritten control; E: engineering check. On/off denote the signal mechanism. The horizon column is simulation time; the other columns are real UTC intervals. A dash means no such arm was attempted under that label. Run Horizon (sim s) On (elapsed s) Off (elapsed s) P01 90.0 90.067 90.106 P02 90.0 90.110 90.064 P03 90.0 90.107 90.056 P04 90.0 90.091 90.106 P05 90.0 90.183 90.122 P06 90.0 90.051 90.062 H01 90.0 0.663 0.521 H02 90.0 0.478 0.477 H03 90.0 1.179 0.678 H04 90.0 1.213 0.572 H05 90.0 0.542 0.539 H06 90.0 0.511 0.527 Pilot1 30.0 30.172 — Pilot2 30.0 30.074 — E1 2.0 2.062 — E2 2.0 — 2.060 The retained synthetic suite and clean reproduction took 3.868 and 3.173 s between recorded UTC captures. Each includes 52 runtime executions, 72 signal cells, 192 inheritance executions, and 32 immigration executions. Native Date captures remain real despite virtual Date.now/performance.now overrides; they exclude earlier startup and later cleanup. Individual fixture elapsed times were unrecorded. Runtime horizons of 4.137 simulation seconds (pause: 0.290 s) are not elapsed-runtime benchmarks. The launcher failure and incomplete harness execution (Appendix E) have unknown elapsed times, not zero cost. Original run directories occupy 894,474,997 logical bytes; complete and failed fixture directories occupy 141,942,445 and 22,891 bytes: 1,036,440,333 bytes (1.036 decimal GB) total. Dependencies, extra analysis/build files, and temporary space are additional. The frozen source archive adds 3,279,256 compressed bytes. These are file lengths, not peak disk requirements. Per-run intervals, size definitions, and source-record hashes accompany evidence/compute-resources.json. Provider hardware and energy are undisclosed; API access does not require their disclosure. Offline reanalysis uses saved responses; exact future API behavior is not guaranteed. CPU/GPU utilization, peak resident memory/storage, and aggregate development, earlier exploratory work, testing, typesetting, and assistant-service compute were unmeasured. No complete project-wide compute total or hardware-efficiency result is claimed.