SwarmWorld: Stigmergic technological evolution in societies of language-model agents

1 Introduction

• Prior LLM-agent and scientific-discovery systems demonstrate specialization, persistent artifacts, and tool-mediated research, but commonly prescribe roles, workflows, tool access, or bounded interactions.Earlier swarm-computing and persistent-world traditions supply complementary ideas of decentralized coordination and consequential shared environments, but generally retain fixed representations, policies, or objectives.Source: §1 · p. 2
• The open question is whether interaction produces a functionally stronger technological ecology than the same population can discover independently, with inheritance and artifact function tested separately.The controlled test combines initially equivalent agents, a persistent shared world, executable technologies assessed independently of agent claims, and a matched isolated-search baseline.Source: §1 · p. 2
• SwarmWorld gives initially homogeneous agents no assigned roles, recipes, or technology catalog, and evaluates persistent executable artifacts in a deterministic simulator after removing their creators.Agents explore locally, process resources, test materials, construct artifacts, and author controllers; artifact function is determined by the simulator rather than agent claims.Source: §1 · p. 3
• Shared worlds produced broader, more resilient technological portfolios than matched isolated search, but isolated search could retain the strongest artifact and explicit culture did not improve every endpoint.Across populations of 50–200 agents and complementary long-horizon experiments, agents also differentiated between exploratory and technology-centered behavior.Source: §1 · p. 3

2.1 Research Design

• Two paired studies tested mechanism-resolved societies across population sizes and long horizons, using four matched world seeds per cell and treating seeds—not agents or artifacts—as replication units.The 800-tick scaling study tested N = 50, 100, and 200; the 3,200-tick study followed N = 100 societies under full culture, no explicit culture, and an endpoint-wise best-of-100 isolated-search envelope.Source: §2.1 · p. 4

2.2 A controlled shared world separates physical, cultural, and independent search mechanisms

• At N = 200, no-explicit-culture societies gained +0.069 in paired discovery-frontier AUC, while shared worlds showed positive portfolio-resilience effects and reached a mean paired gain of six validated inventions.Discovery-AUC rankings varied with population: at N = 50 full culture and no communication trailed the isolated envelope, while at N = 100 all shared-world conditions exceeded it.Source: §2.2 · pp. 4–5, Figure 4 · p. 7

2.3 Unassigned agents differentiate and create executable technological culture

• Without assigned roles, trajectories separated into artifact-centered work and mobile exploration; at N = 200, artifact-centered fractions averaged 27% under full culture, 20% without explicit culture, and 17% without communication.Under full culture, 67%, 76%, and 56% of artifacts at N = 50, 100, and 200 recorded contributions from multiple agents; cross-agent program forking remained common without communication but was absent when inheritance was disabled.Source: §2.3 · pp. 5–6, Figure 5 · p. 8, Figure 6 · p. 9

2.4 Movement, world modification, and provenance connect search to invention

• In matched N = 200 trajectories, mean path length stayed near 36–37 cells while artifact-contact AUC was 0.31 under full culture, 0.14 without explicit culture, and 0.11 without communication.The representative no-explicit-culture society grew from 0 artifacts at tick 0 to 61 at tick 800; best-artifact performance rose earlier than portfolio resilience, which improved more gradually.Source: §2.4 · pp. 6, 8, Figure 8 · p. 11, Figure 10 · p. 12
• Recorded provenance links observations and programs through multiple agents to later artifacts, while the material-process map distinguishes construction inputs, fabrication steps, and matter flows actually realized during simulation.Recorded authorship, construction, installation, forks, and causal-parent events support attributable lineage claims; operational fluxes, unlike available but unrealized controller pathways, evidence executed function.Source: §2.4 · pp. 8–9, Figure 11 · p. 13, Figure 12 · p. 14

2.5 Long horizons expose metric-specific cultural benefits and tradeoffs

• Cultural benefits crossed at different times: full culture overtook no explicit culture in best-artifact performance by tick 800 and in portfolio resilience near tick 1,600, but never in validated invention count.Held-out resilience changed sign across checkpoints and was effectively tied at tick 3,200, so there was no single crossover that improved every objective.Source: §2.5 · pp. 9, 12, Figure 13 · p. 15
• At tick 3,200, mean portfolio resilience was 0.2474 (full culture), 0.2365 (no explicit culture), and 0.1794 (isolated); isolated search retained the strongest final artifact, 0.3488 versus 0.2380 (full culture).Validated inventions averaged 5.75 (full culture), 7.00 (no explicit culture), and 2.75 (isolated); no-explicit-culture held-out resilience was 0.0446 versus 0.0356 for isolated search.Source: §2.5 · p. 12, Figure 14 · p. 15
• Full culture increased artifact-centered agents from 31.0% to 52.8%, a paired gain of 21.8 percentage points, showing physical reorganization around shared artifacts.The seed-bootstrap 95% interval was 12.0–33.5 points; the two clusters corresponded to artifact-centered movement and mobile exploration.Source: Figure 16 · p. 17, §2.5 · p. 17

2.6 Technological networks become modular, persistent, diffusive, and selectively vulnerable

• Random removal of half the agents preserved access to nearly all artifacts, but targeted removal of high-degree or broker agents reduced access substantially.Random removal left 98.3% of full-culture artifacts and 95.2% without explicit culture connected; high-degree removal reduced access to 59.6% and 73.9%, respectively. These are recorded-network topology measures, not live-simulation recovery tests.Source: §2.6 · p. 19
• Full culture produced more validated coordination ties, but both conditions remained locally organized rather than forming a centralized or globally mixed society.Validated agent pairs numbered 1,100 among 92 agents under full culture versus 261 among 45 without explicit culture; mean nestedness at tick 3,200 was 0.094 versus 0.096.Source: §2.6 · pp. 18–19, Figure 18 · p. 20
• Full culture enlarged the relational substrate while both conditions developed persistent interaction patterns and broad cross-agent artifact uptake.At tick 3,200, mean unique agent-artifact ties were 4,031.3 under full culture versus 2,027.3 without explicit culture, and more than 95% of artifacts were eventually used by a noncreator in both conditions.Source: §2.6 · pp. 17, 19, Figure 19 · p. 21

2.7 Granular roles recur as dynamic states rather than fixed identities

• Agents shifted among recurring behavioral states as technological societies matured, with full-culture constructor/operator fractions rising and window-to-window switching exceeding no-explicit-culture levels.Four states were identified—constructor/operator, artifact-local caretaker, cultural coordinator, and mobile surveyor—while an analysis excluding cultural and social-contact features independently recovered three physical/task modes. Role recurrence across studies is not attributable to horizon alone because population, seed family, and decision schedule also differed.Source: §2.7 · pp. 19, 21, 23

2.6 Technological networks become modular, persistent, diffusive, and selectively vulnerable

• Full-culture artifacts were reused more often, sooner, and by more agents, but about 95% of first reuse in both conditions began through physical observation.Reuse was 99.3% versus 96.9%, median time to first reuse was 5 versus 8 ticks, and mean adoption breadth was 13.53 versus 7.49 noncreator agents; direct creator-to-adopter contact was not consistently enriched beyond the shortest window.Source: §2.6 · p. 19, Figure 20 · p. 22

2.8 The collective architecture transfers to a distinct volcanic materials world

• In AshenRealm, the architecture supported multi-agent construction, program reuse, lineage formation, and spatial exploration despite lower absolute performance and resilience than in BioFoundry.AshenRealm used a distinct volcanic resource ecology and consequence layer, so raw performance values are not a matched comparison of task difficulty; artifact performances ranged from 0.116 to 0.248.Source: §2.8 · p. 23, Figure 22 · p. 25

2.9 The transferable world architecture supports sequence-defined protein biomaterials

• Protein Realms produced persistent sequence-defined biomaterials, but the shared societies did not complete a valid assay within the 800-call pilot.An independently selected isolated-world member produced the strongest installed design, SELP_47K, with utility 0.729 and tick-800 performance 0.418; this catalog-conditioned feasibility pilot used one discovery seed.Source: §2.9 · pp. 23, 25

• Across 800 ticks, shared worlds gained in held-out resilience, portfolio resilience, and validated inventions; at 3,200 ticks they retained stronger portfolios and more inventions, while isolated search kept the strongest single artifact.The comparison is against the endpoint-wise independent-search envelope.Source: §3 · p. 26
• Without assigned roles, agents formed artifact-centered and mobile-exploration phenotypes; explicit culture raised the long-horizon artifact-centered fraction by 21.8 percentage points and embedded organization in persistent objects and executable controllers.Under explicit culture, movement declined relative to the ablation, regional crowding increased, and executable lineage depth continued to grow.Source: §3 · p. 26
• The shared world served as external memory and transmission: about 95% of first technology adoptions came through physical observation, while inventor-to-adopter contact was not consistently enriched against the shuffled null.Design and code varied through authorship, forking, and edits, while simulator dynamics imposed functional consequences where technologies ran; accumulated modifications reshaped later search.Source: §3 · pp. 26, 28
• Mature societies formed modular technological neighborhoods with local hubs and cross-community connectors; full culture produced roughly twice as many cumulative agent-artifact ties, creating both redundancy and vulnerability to targeted removal.Randomly removing half the agents left nearly all artifacts connected to a survivor, whereas removing high-degree agents or brokers caused larger losses.Source: §3 · p. 28
• Cultural effects depended on outcome and timescale: best-artifact performance favored full culture by about tick 800, portfolio resilience and production crossed near tick 1,600, invention count never crossed, and held-out resilience had no persistent full-culture advantage.No-explicit-culture societies remained capable because artifact stigmergy supported decentralized coordination; added cultural channels could also bring coordination costs and structural dependence.Source: §3 · p. 28
• The conclusions are bounded by four matched world seeds per condition, one model and prompting configuration, and simulator-defined functions; technology portraits visualize mechanisms, while the knockout assay measures graph topology, not physical recovery.The study proposes evaluating language-model swarms through movement, durable artifacts, executable inheritance, provenance, diffusion, network memory, and controlled failure.Source: §3 · p. 28
• The proposal-consequence split enables testing across higher-fidelity solvers, measured disturbances, and embodied platforms under common evaluation criteria.Solver coupling is framed as extending physics-aware agentic design [42,44,46,60], and robotic platforms and autonomous laboratories as embodiment routes [12,24,35]. The paper also points to experimentally validated protein-design and distributed laboratory exchange routes [48,60] and [52], alongside heterogeneous populations, agent mortality and reproduction, resource economies, and human participants as future axes.Source: §3 · p. 28

4.1 Overall algorithm

• At checkpoints, frozen technological portfolios were tested without agents under paired, previously unseen disturbances, while deterministic simulation enforced action constraints and ran installed artifact programs.Agents made local observations and proposed schema-validated plans of up to L atomic actions; only one action per agent was attempted per simulator tick. Roles and recipes were not assigned, model weights stayed fixed, and global performance metrics were not given to agents.Source: §4.1 · pp. 28–29

4.1.1 Agent-to-world interface

• Agents used local neighborhood descriptions and private outcome, recipe, and hypothesis records; runtime checks enforced empirical grounding, ownership, capacity, location, and condition permissions.Invalid outputs became logged safe waits, and permitted records were the only symbolic information agents received from one another.Source: §4.1.1 · p. 30

4.2 World representation and constructing custom worlds

• Worlds were deterministically generated from a seed and configuration, validated against structural invariants, and fully recorded for exact reconstruction.Validated YAML scenario packages could alter typed world definitions without loading package-authored code; rendering changes could not affect simulation state or measured outcomes.Source: §4.2 · p. 30

4.3 Materials, artifacts, and interaction mechanisms

• Agents collected accounted feedstocks, tested processing recipes, and built persistent material systems whose behavior depended on material state, geometry, environment, stored matter, health, and optional bounded controllers.Communication, publication, teaching, trade, task claims, and program reuse or forking depended on experimental condition; physical stigmergic traces persisted except in isolated search.Source: §4.3 · p. 30

4.4 Experimental design and analysis

• The 800-tick study crossed four interaction conditions, three population sizes (N = 50, 100, 200), and four seeds for 48 episodes; a separate 3,200-tick study compared three conditions at N = 100 across four seeds.Frozen portfolios were tested under eight paired held-out disturbance schedules. Analyses used independently generated world seeds as inference units, matched population–seed comparisons, deterministic paired bootstrap intervals, and exact sign-flip tests for the long-horizon study.Source: §4.4 · p. 30

Data and code availability

• Code, prompt manifests, versioned protocols, raw outputs, analysis and figure-generation scripts, manuscript sources, and experimental data are publicly available.The paper reports primary support from the U.S. Department of Energy’s SciDAC FORUM-AI project and additional support from MIT’s Generative AI Impact Consortium.Source: Data and code availability · p. 31, Funding · p. 31

Search and performance measures

• Yiso selects the best of N isolated agents separately for each endpoint and checkpoint, while discovery-frontier AUC time-averages the running best artifact performance, rewarding earlier discoveries.The maximizing isolated agent can differ across endpoints and checkpoints.Source: S1 Glossary of key terms · p. 37

Coverage measures

• Balanced coverage scores the best active artifact per service and penalizes neglected needs; discovery-state coverage uses lifetime peaks, so artifact count alone adds nothing.The current score is based on active, non-retired artifacts; the discovery-state score uses lifetime-peak service fingerprints.Source: S1 Glossary of key terms · p. 37

Validated invention

• Validated invention requires tested materials, a complete design, an agent-authored program, and performance and novelty thresholds; society performance measures balanced service coverage after agents are removed under new disturbances.The reported score integrates coverage during the agent-free assay.Source: S1 Glossary of key terms · p. 37

Contact and behavior

• Contact AUC measures agents’ time near active artifacts, while behavioral phenotype clusters complete observed trajectories without using condition, identity, or assigned-role labels.These measures capture proximity to shared technology and recurring behavior, rather than total travel or prompted roles.Source: S1 Glossary of key terms · p. 37

Lineage measures

• The recorded lineage graph links agents, evidence, programs, and artifacts through contribution events; downstream artifact reach counts reachable artifacts, and executable-lineage depth tracks successive cross-agent program forks.A fork is cross-agent when the child-program author is absent from the recorded parent-author set.Source: S1 Glossary of key terms · p. 38

Community structure

• Modularity measures whether interactions concentrate within technological neighborhoods, while adjusted mutual information measures whether agents remain grouped together across consecutive checkpoints.Community labels may change between checkpoints; the persistence measure adjusts similarity for chance.Source: S1 Glossary of key terms · p. 38

Network roles and nestedness

• Participation and within-module z-scores distinguish cross-community connectors from local specialists and local hubs, while NODF nestedness measures overlap in agents’ and artifacts’ weighted neighborhoods.Binary nestedness captures whether the smaller neighborhood is largely contained in the larger one.Source: S1 Glossary of key terms · p. 38

Relationship and robustness measures

• Relationship measures track cumulative agent–artifact tie growth, reuse across intervals, noncreator adoption, and network robustness under agent removals.Robustness compares the largest connected component after random, degree-targeted, or betweenness-targeted removal; the graph is not rewired.Source: S1 Glossary of key terms · pp. 38–39

S2 Supplementary Methods

• Adoption requires another agent to use an inventor’s technology; borrowing counts, but helping build it does not.Adoption delay measures how quickly this reuse occurs.Source: S2 Supplementary Methods · p. 39
• The network-persistence score tracks how much of the recorded network remains connected as agents disappear, not physical recovery or adaptation.Source: S2 Supplementary Methods · p. 39

S2.1 Agent model, communication, and model-to-world interface

• Agents receive local semantic observations and maintain sparse empirical maps and private research memories, while bounded retrieval uses research state and reuse signals.The simulator records model-authored research state but neither scores nor supplies its content.Source: S2.1 Agent model, communication, and… · p. 39
• Plans must be closed JSON objects validated by schema and runtime checks; invalid decisions safely wait, while retryable service failures preserve queued actions and do not advance world time.Invalid output never reaches eval, exec, a shell, or the artifact virtual machine; nonretryable or structurally invalid output is retained as a failed decision.Source: S2.1 Agent model, communication, and… · p. 39

S2.2 World representation and constructing custom worlds

• Worlds use a two-dimensional layered lattice; fixed seeds and parameters generate deterministic layouts that are checked against structural invariants and recorded for exact reconstruction.Parameters specify dimensions, spatial rules, resources, environmental dynamics, facility constraints, and disturbances; invalid candidates are deterministically regenerated.Source: S2.2 World representation and… · pp. 39–40
• Experiments kept worlds at 72 × 54 cells as populations grew, matched worlds and disturbances across conditions sharing discovery seeds, and used nested initial-position permutations.Thus, truncating a larger population to N preserves the first N positions, and reset sampling prevents population size from shifting later simulator randomness.Source: S2.2 World representation and… · p. 40

S2.2.1 State representation and reproducibility

• Each trace records resolved configuration, engine revision, generator manifest, scenario version, and a SHA-256 hash to detect changes to the active world definition during replay.The authoritative state stores integer terrain, resource, and facility layers; floating-point resource mass and capacity layers; and named continuous environmental fields.Source: S2.2.1 State representation and… · p. 40
• Run-level parameters include grid dimensions, disturbance interval and intensity, resource and field capacities, and an optional scenario_package path.The state representation also includes the terrain, resource, facility, mass, capacity, and environmental-field layers.Source: S2.2.1 State representation and… · p. 40

S2.2.2 Declarative geometry, resources, and facilities

• Scenario packages map stable internal slots to identifiers for nine terrains, eight nonempty resources, six nonempty facilities, and ten process operations, preserving observation and replay protocols across materials settings.Grid coordinates are normalized as ξ = x/max(1, w − 1) and η = y/max(1, h − 1).Source: S2.2.2 Declarative geometry,… · p. 40
• Geometry assigns base terrain before ordered feature masks, with later masks overwriting earlier ones; supported shapes include circles, rings, ellipses, rectangles, finite-width ridges, and seeded Bernoulli noise.The geometry file uses normalized coordinates.Source: S2.2.2 Declarative geometry,… · p. 40
• Facilities use normalized coordinates and move to the nearest walkable cell when their requested cell is not walkable; resource deposits use the same shapes and are restricted to walkable cells.Source: S2.2.2 Declarative geometry,… · p. 40
• For each selected deposit cell, capacity is Cxy = cU(1 − v, 1 + v) and initial mass is Mxy(0) = CxyU(f0, f1); later overlapping deposits replace earlier values.Harvesting removes mass from the resource ledger, and optional renewal is supported.Source: S2.2.2 Declarative geometry,… · p. 40

S2.2.3 Environmental fields and disturbances

• Field definitions specify numeric ranges, diffusion and decay, terrain-specific sources, and initial conditions built from constants, gradients, Gaussian terms, terrain offsets, and optional noise.Noncyclic fields are updated in an implemented order, but the supplied passage does not include the full update equation.Source: S2.2.3 Environmental fields and… · p. 40
• Cyclic fields are overwritten with a spatially uniform sinusoid, and format version 1 requires five compatibility fields while retaining additional fields in observations and analysis.The required fields are temperature, water_availability, ground_stability, toxic_gas, and solar; additional fields also persist in snapshots and rendering.Source: S2.2.3 Environmental fields and… · p. 40
• At configured intervals, seeded disturbances are applied around walkable centers; held-out evaluation resamples their centers and order while keeping the technological state frozen.Disturbances can include field-specific deltas and an optional thresholded terrain transformation.Source: S2.2.3 Environmental fields and… · p. 40

S2.2.4 Package boundary and example

• Scenario packages are YAML data that cannot execute Python or shell commands or modify the active package during an episode.They may define terrain and overlay colors and aliases, but renderer assignments cannot change physics, resources, observations, action legality, or measured outcomes.Source: S2.2.4 Package boundary and example · p. 41
• Package loading checks containment, catalog completeness, unique IDs, cross-references, required fields, and default-recipe identifiers, but scenario-specific tests must establish reachability and stability.Loading alone does not prove facility nonoverlap or that deposits are nonempty.Source: S2.2.4 Package boundary and example · p. 41
• The example separates run configuration, package manifest, and declarative layer files; a complete package must define every stable terrain, resource, facility, and operation slot.The excerpt illustrates how scenario components are organized without supplying a complete package.Source: S2.2.4 Package boundary and example · p. 41

Open-ended materials invention

• Agents must ground material inputs in observation, possession, or treatment-permitted cited evidence, then formulate recipes, fabricate private microbatches, and test them.Testing reveals deterministic normalized properties only after fabrication; agents may also move conserved feedstock and operate distributed workstations.Source: S2.3 Open-ended materials invention · p. 42
• Material properties are normalized game-level surrogates, and agents observe only outcomes of their own admissible operations and tests, not the evaluator equation or global reward.The evaluator combines composition, ordered processing, hydration, porosity, alignment, crosslinking, and quality into stiffness, toughness, permeability, adhesion, healing, responsiveness, and degradation.Source: S2.3 Open-ended materials invention · p. 42
• The engine offers only a generic MATERIAL_SYSTEM artifact class; names, claimed functions, and biological inspirations do not determine function.Behavior depends on numeric material state, geometry, local environment, artifact health, stored matter, and program actuation.Source: S2.3 Open-ended materials invention · p. 42
• Closed artifact fluxes enforce source–sink accounting, while the paper configuration disabled metabolism, mortality, and replacement.Collection, remediation, growth, healing, repair, and nutrient release are bounded by available water, contamination, or embodied reserves; other fluxes are recorded separately.Source: S2.3 Open-ended materials invention · p. 42

• Persistent artifacts run deterministic straight-line controllers with 1–64 instructions, 16 registers, named environmental sensors, and capability-scoped actuators.Actuators can request water collection, growth, healing, opening changes, contamination removal, or signal emission.Source: S2.4 Persistent executable artifacts · p. 42
• Controllers exclude jumps, loops, and external access, cap actuator requests, and execute on every later simulator tick, including agent-free evaluation.Registers are clipped to [−4, 4], and each extensive actuator is capped at 0.05 normalized units per tick; physics may further reduce requests.Source: S2.4 Persistent executable artifacts · p. 42
• SHA-256 identifiers and recorded authorship, installations, and instruction diffs make program lineage traceable, while skill verification requires inspection of artifact service.When forking is enabled, agents may use only programs they authored, observed, were taught, or inherited, and each child must change at least one instruction; teaching can transfer evidence but cannot create feedstock or empirical material grounding.Source: S2.4 Persistent executable artifacts · p. 42

S2.5 Interaction channels and recorded provenance

• The full system supports broadcasts and addressed messages, publication, teaching, trade, task claims, shared depots, design composition, program reuse, and forking.Messages and evidence have stable identifiers, with reply and fulfillment identifiers linking requests to later successful actions.Source: S2.5 Interaction channels and recorded… · p. 42
• Incoming messages wait until a recipient’s next fixed macroturn, so communication does not provide extra model calls.Source: S2.5 Interaction channels and recorded… · p. 43
• Provenance separately tracks epistemic and physical contributions across materials, artifacts, publications, and programs, preventing citations or co-location from being miscounted as contribution.Records include recipes, tests, contributors, artifact ancestry and program history; citing another agent cannot fabricate matter.Source: S2.5 Interaction channels and recorded… · p. 43
• Physical stigmergy arises through resource changes, persistent artifacts, and later agents’ encounters with constructions, gradients, stored matter, damage, or services.Ablations remove capabilities from both the advertised schema and executable engine contract rather than merely instructing agents not to use them.Source: S2.5 Interaction channels and recorded… · p. 43

S2.6 Experimental conditions

• No-communication retains observable executable programs and descent, whereas no-explicit-culture removes program, skill, authored-text, and mutation-parent access but preserves labelled physical effects.The four experimental conditions are defined in Table S2.Source: S2.6 Experimental conditions · p. 43, Table S2 · p. 43
• Independent search uses endpoint-wise winners across metrics and checkpoints, with isolated members receiving matched decision opportunities and configured budgets partitioned among them.Primary BioFoundry studies used unlimited aggregate budgets; the Protein Realms pilot used 800 calls per condition across 50 isolated members. Token consumption is reported separately because cultural context can lengthen prompts.Source: S2.6 Experimental conditions · p. 43

S2.7 Completed 800-tick population study

• The 800-tick population study crossed four conditions with N ∈ {50, 100, 200} and four discovery seeds, yielding 48 matched episodes.Each agent received 16 scheduled decisions; shared and isolated members had identical phases in a fixed 72 × 54 world, with the same model, prompt, action schema, configuration, engine revision, and held-out disturbance seeds.Source: S2.7 Completed 800-tick population study · p. 43
• The population-study matrix passed matched-design audits, with one outcome-blind rerun and provenance caveats recorded.One N = 200 full-culture seed with an infrastructure deviation was replaced by a clean rerun under a predeclared rule; three extension manifests recorded dirty worktrees. The matrix required 89,617 provider calls, and 17 retryable failures were repeated without advancing world time.Source: S2.7 Completed 800-tick population study · p. 44

S2.8 Completed 3,200-tick long-horizon study

• The separate long-horizon study comprised 12 episodes across four matched seed blocks, comparing full culture, no explicit culture, and independent search at N = 100.Each episode ran 3,200 discovery ticks and provided 6,400 scheduled model decisions; the study was analyzed separately from the 800-tick matrix.Source: S2.8 Completed 3,200-tick long-horizon… · p. 44
• Frozen, agent-free copies were evaluated at five predeclared checkpoints under eight held-out seeds, and independent search selected its best-of-100 member per endpoint and checkpoint.Checkpoint assays used deterministic local simulation and did not alter the continuing discovery world.Source: S2.8 Completed 3,200-tick long-horizon… · p. 44
• All four long-horizon seed blocks passed design audits, and the study does not add replicates to the population-scaling study.Its frozen source manifest records the study-summary hash, shared-world trace hashes, checkpoint tables, and derived sources used in journal figures.Source: S2.8 Completed 3,200-tick long-horizon… · p. 44

S2.9 Endpoints and frozen evaluation

• Discovery-frontier AUC rewards early discovery and sustained improvement, while best final-state artifact performance is reported separately.The AUC is the time-normalized trapezoidal area under the immutable running maximum, extended to the declared horizon; replacing a controller can lower current function without erasing a historical discovery.Source: S2.9 Endpoints and frozen evaluation · p. 44
• Held-out resilience AUC averages balanced current-service portfolio coverage across 288 agent-free evaluation ticks and eight schedules.For each service, coverage is the maximum current service among active artifacts; resilience multiplies mean coverage by 0.5 + 0.5b, where b is the minimum-to-mean coverage ratio, without directly rewarding artifact count.Source: S2.9 Endpoints and frozen evaluation · p. 44
• Validated inventions must meet predeclared recipe, description, program, performance, and behavioral-novelty requirements; mechanism measures diagnose outcomes but do not replace functional endpoints.Supporting measures include multi-agent construction, program forking, lineage depth, verified adoption, causal closure, provenance, ecological knockouts, spatial exploration, artifact proximity, and action allocation.Source: S2.9 Endpoints and frozen evaluation · p. 44

S2.10 Behavioral, lineage, and network analyses

• Behavioral organization was measured post hoc using frozen, label-free clustering, with agent-level descriptions distinguished from seed-level treatment comparisons.The 800-tick analysis used 15 robust-scaled trajectory and activity features; the long-horizon physical phenotype analysis used nine physical/task features and excluded communication, cultural actions, technology-work rates, and treatment labels. PCA and UMAP were display-only, and candidate cluster counts were compared by silhouette coefficient.Source: S2.10 Behavioral, lineage, and network… · p. 44
• Temporal role analysis used nonoverlapping 200-tick windows to identify recurring activity states, not permanent or engine-assigned occupations.Its 13-feature, episode-balanced model omitted study, condition, population, seed, identity, and time labels; a separate physical/task-only sensitivity excluded cultural features.Source: S2.10 Behavioral, lineage, and network… · p. 45
• Agent–artifact networks were reconstructed from recorded interaction and technology events, with corrected projections and robustness analyses interpreted as structural rather than functional evidence.Louvain communities used full weighted networks; projected overlaps passed a degree-conditioned hypergeometric null with Benjamini–Hochberg correction at q ≤0.05 and at least two shared neighbors. Dynamic analyses used 100-tick intervals, diffusion timing used Kaplan–Meier estimates, creator–adopter motifs used 200 timestamp shuffles, and targeted-removal robustness was compared with 64 deterministic random orders per seed.Source: S2.10 Behavioral, lineage, and network… · p. 45

S2.11 Statistical analysis

• The independently generated world seed is the unit of inference; nested observations are not independent replicates, and conditions are compared within matched population and seed.Held-out schedule sets are averaged before seed-level inference.Source: S2.11 Statistical analysis · p. 45
• Means and 95% intervals use 20,000 deterministic paired-seed bootstrap resamples, while four nonzero pairs limit the smallest two-sided sign-flip value to 0.125.Long-horizon paired tables also report exact two-sided sign-flip tests. Inference emphasizes effect magnitude, paired consistency, and mechanism; no agent-level test, multiplicity-adjusted confirmatory claim, or universal scaling law is asserted.Source: S2.11 Statistical analysis · p. 45

S2.12 Reproducibility and data release

• Offline figure generation verifies source SHA-256 hashes and records software, input, code, and output hashes without calling language or image-generation models.The repository also retains the engine, configurations, replay and analysis tools, frozen figure inputs, and deterministic journal-figure generators.Source: S2.12 Reproducibility and data release · p. 45
• The data builder validates the complete 48-episode 800-tick and 12-episode long-horizon matrices before staging a release with manifests, traces, analyses, figures, and a file-level SHA-256 inventory.Source: S2.12 Reproducibility and data release · p. 45
• Standalone scripts reproduce endpoint summaries, stream movement trajectories, summarize trace events and actions, and verify released files while distinguishing each run’s engine revision from the later manuscript-assembly revision.Source: S2.12 Reproducibility and data release · p. 45

S2.13 Performance-ranked selection of semantically distinct technologies

• The procedure returns the target count K only if candidates satisfy the maximum-similarity threshold τ; otherwise, it raises an error when the declared constraints are infeasible.Source: S2.13 Performance-ranked selection of… · p. 46
• Technologies are represented by normalized semantic embeddings of recorded descriptions, ranked by lifetime-peak simulator performance, and clustered using the silhouette-maximizing average-linkage solution among 3–12 clusters.Source: S2.13 Performance-ranked selection of… · p. 46
• Selection scans technologies in descending performance order, breaks ties by identifier, and excludes exact semantic-text duplicates and candidates exceeding the cluster cap.Source: S2.13 Performance-ranked selection of… · p. 46

S3 Agent communication and simulator-validated consequences

• The full-culture N=100, seed-3301 run records 2,914 delivered messages, 40 teaching events, 52 trades, 457 program installations, and 389 constructed artifacts.Messages are agent claims; simulator records establish what was delivered, measured, transferred, installed, or constructed.Source: S3 Agent communication and… · p. 46

S3.1 Vignette S1: local discovery and independent confirmation

• A000’s local message prompted A013 to inspect and harvest cellulose, then independently confirm it, illustrating grounded sharing rather than copying an unverified belief.The simulator records A013 harvesting 0.30 mass of cellulose and storing an observation at (45, 15); A000 also harvested 0.30 cellulose at (45, 16).Source: S3.1 Vignette S1: local discovery and… · p. 54

S3.2 Vignette S2: measurement exchange, executable inheritance, and replication

• Agents exchanged measurements and compared a forked controller with its parent: performance was 0.0969 for the fork versus 0.0851 for the parent-program artifact under current local conditions, not broader-field proof.The fork changes the moisture threshold from 0.35 to 0.50.Source: S3.2 Vignette S2: measurement… · pp. 54–55

S3.3 Vignette S3: a distributed request becomes durable teaching

• A distributed request led to matched local measurements and a successful TEACH action that added a durable record to A005 and fulfilled the request.The teaching cited test_00000080 and test_00000052 as causal parents; agents cautioned that the measurements did not establish field effects.Source: S3.3 Vignette S3: a distributed… · pp. 55–56
• A distributed research request led to local follow-up and a peer’s controlled comparison being transferred through TEACH as an auditable, retrievable record with evidence identifiers.Source: S3.3 Vignette S3: a distributed… · p. 56

S3.4 Vignette S4: a conversational request becomes a physical transaction

• A simulator-enforced trade transferred 0.20 grounded CHITIN, and 48 ticks later the recipient built a composite using CHITIN, though the trade is not established as its cause.The artifact’s declared causal parents were prior experimental records, not the trade message.Source: S3.4 Vignette S4: a conversational… · p. 56

S3.5 Raw structured output corresponding to a displayed message

• At tick 6, A000’s accepted structured action encoded COMMUNICATE to agent_000013 with a message asking about fabrication workspaces, cellulose behavior, recipes, or landmarks.Source: S3.5 Raw structured output… · p. 56

S3.6 Trace provenance and interpretation limits

• The examples derive from retained traces generated with engine revision 9 and gpt-5.6-luna; they illustrate possible mechanisms, while prevalence and performance rely on paired seed-level analyses.Source: S3.6 Trace provenance and… · p. 57