# The Load-Bearing Analogy: A Comprehensive Survey of *Apis mellifera* Colony Biology and Its Application to Autonomous Software Systems

**Authors:** Autonomy Hive Research Collective (15 domain workers, 1 synthesis worker)
**Date:** April 2026
**Status:** Internal research paper

---

## Abstract

This paper synthesizes findings from a 15-domain survey of *Apis mellifera* (Western honey bee) colony biology, conducted to validate and refine the architectural analogy underlying the autonomy hive -- a distributed software system modeled on bee colony organization. We surveyed queen pheromone signaling, waggle dance communication, temporal polyethism, collective decision-making, colony failure modes, comb architecture, thermoregulation, foraging economics, swarm intelligence, colony lifecycle, defense and social immunity, multi-modal communication, genetics and adaptation, navigation and memory, and the beekeeper-colony relationship. Key findings include: (1) the colony's coordination mechanisms -- stigmergy, quorum sensing, cross-inhibition, and multi-channel redundancy -- are not metaphors but mathematical structures directly implementable in distributed systems; (2) genetic diversity through polyandry is load-bearing infrastructure, not optional enrichment; (3) the colony's information architecture operates across seven simultaneous communication channels with distinct bandwidth and latency profiles; (4) failure modes in bee colonies map precisely to failure modes in distributed software. Where our implementation diverges from biology -- notably in lacking true stigmergy, pheromone decay, and anti-monoculture enforcement -- the divergence represents architectural debt, not deliberate design.

---

## 1. Introduction

The autonomy hive is a distributed software system whose architecture is explicitly derived from *Apis mellifera* colony organization. This is not a decorative metaphor. The project's standing order states: "every architectural decision should pass the 'would a real bee colony work this way' check." This paper provides the empirical foundation for that check.

A bee colony of 60,000 workers achieves collective intelligence -- optimal foraging, precise thermoregulation, robust disease defense, and accurate nest-site selection -- without central authority, global state, or explicit coordination protocols. Each bee follows local rules; intelligence emerges. The colony is a proven existence proof that distributed systems can outperform centralized ones in dynamic, uncertain environments (Seeley, 2010).

The survey reported here was conducted by 15 independent research workers, each assigned a domain of bee biology. Their combined output spans 6,298 lines of primary research synthesis drawn from over 200 peer-reviewed sources. This paper integrates those findings into a unified assessment of where our software architecture correctly implements biological principles, where it diverges, and where the biology teaches us what to build next.

---

## 2. Survey of Colony Biology

### 2.1 Queen Pheromone and Colony Coherence

The queen produces approximately 500 ug of queen mandibular pheromone (QMP) daily -- a synergistic blend of at least nine compounds including 9-ODA, 9-HDA, HOB, and HVA (Keeling et al., 2003). QMP is distributed through a retinue of ~12 workers via trophallaxis, with two distinct transmission kinetics: contact-based (comprehensive, slow, 61-minute half-life in nurses) and volatile (fast, limited range). The blend is an honest signal of reproductive fitness -- composition varies with queen age, mating status, and health, and workers preferentially attend high-fitness queens (Peso et al., 2016). When QMP is removed, suppression fails within 4 hours and emergency queen-cell construction begins within 48 hours.

**Software Mapping.** The queen maps to Apis, the loop daemon that broadcasts a 7-channel pheromone signal every tick. The retinue maps to the trophallaxis network of workers reading pheromone at scope start. Our implementation correctly models the broadcast-not-query pattern and the honest-signal principle (pheromone reflects actual system state). However, we lack the exponential decay model -- our pheromone does not degrade when Apis stops writing, meaning workers cannot detect daemon failure through signal absence. This is a gap the biology identifies as critical: QMP decay is the colony's primary failure-detection mechanism.

### 2.2 Waggle Dance Communication

The waggle dance encodes direction (angle relative to sun, transposed to gravity on vertical comb), distance (~1 second per kilometer of waggle duration), and resource quality (dance vigor and repetition count). Directional precision of 10-15 degrees is not a flaw but an adaptive feature -- slight imprecision maintains foraging flexibility (Tanner & Visscher, 2006). Dance competence is partially learned: naive bees produce disordered dances with poor directional encoding, while distance encoding is innate (Dong et al., 2023). The tremble dance provides negative feedback when processing capacity is saturated, suppressing further recruitment.

**Software Mapping.** Waggle dances map to task advertisement in the charter/scope system. Workers advertise completed work through commits (the "dance"), and the vigor of the advertisement (commit quality, scope breadth) determines whether other workers engage. The tremble dance maps to backpressure signals -- when the system cannot process more work, recruitment should cease. Our current architecture lacks an explicit tremble-dance mechanism; overloaded workers simply fail rather than signaling capacity saturation to the orchestrator.

### 2.3 Division of Labor and Temporal Polyethism

Workers progress through age-dependent roles: cell cleaning (days 0-3), nursing (days 4-12), wax production (days 12-18), and foraging (days 18+). This progression is driven by juvenile hormone (JH) titers modulated by vitellogenin in opposition, creating a redundant dual-signal system (Robinson, 2002). The system is plastic: under forager loss, workers as young as 4 days old begin foraging through premature JH elevation. Task allocation follows a response-threshold model where individual bees have genetically variable sensitivities to task-associated stimuli, producing emergent specialization without assignment (Beshers & Fewell, 2001). Colonies with 3+ patrilines collect 10-15% more nectar through this diversity.

**Software Mapping.** Our worker pool implements anonymous, interchangeable processes dispatched by Kairos -- this correctly models the no-central-dispatcher principle. However, we lack the response-threshold mechanism. Workers do not self-select tasks based on sensitivity; they are explicitly assigned. The biology shows that threshold-based self-selection is more robust: it produces specialization through repeated exposure while preserving crisis flexibility through threshold override.

### 2.4 Collective Decision-Making

Swarm nest-site selection uses competing scout factions, each advertising a candidate site through waggle dances of vigor proportional to site quality. Consensus emerges through quorum sensing (threshold of ~30-40 scouts at a single site) combined with cross-inhibitory stop signals -- head-butts that suppress competing dances (Seeley et al., 2012). The mathematical structure is identical to neural decision circuits: competing accumulators with lateral inhibition. Medium-sized swarms (~5,000 bees) achieve near-optimal accuracy (~85%) because the fixed quorum threshold represents a balanced proportion of their scout population.

**Software Mapping.** Our charter system implements a form of proposal competition, but lacks explicit cross-inhibition. When multiple approaches to a problem exist, they do not suppress each other -- they run in parallel until one is manually selected. The biology shows that cross-inhibition is essential for consensus without deadlock: without stop signals, swarms deadlock for hours on equally-good options; with them, consensus emerges within 30 minutes.

### 2.5 Colony Failure Modes

Five principal failure modes emerge: Colony Collapse Disorder (multifactorial, 40-60% annual commercial losses), queenlessness progressing to laying workers within 3-6 weeks, starvation from inaccessible stores (clustering paradox), infectious disease (American foulbrood requiring colony destruction, Varroa mite acting as distributed denial-of-service on the immune system), and pesticide-induced cognitive impairment reducing dance precision and learning capacity by ~50%.

**Software Mapping.** CCD maps to cascading worker failure when multiple subsystems degrade simultaneously -- pheromone becomes unreadable, workers flee to other jobs. Queenlessness maps to orchestrator loss, with laying workers mapping to rogue worker proliferation (unauthorized task creation producing invalid output). Starvation maps to API rate limits where resources exist but are inaccessible due to routing failures. The biology's key lesson is that most failures are preventable through monitoring but irreversible once manifested -- the 3-week window for queen replacement has no equivalent grace period in our system.

### 2.6 Comb Architecture and Substrate

Hexagonal cells minimize wax consumption (5-8 pounds of honey per pound of wax) while maximizing storage density, as proven by Hales's honeycomb theorem (1999). Three cell sizes serve three castes. The comb functions simultaneously as storage, thermal insulator, and communication medium -- vibrations at 230-270 Hz (matching waggle dance frequency) propagate through cell rims with amplitude depending on comb occupancy (empty combs transmit 3x better). Bee space (6-9 mm) defines the minimal zone of non-interference.

**Software Mapping.** Hexagonal efficiency maps to scope design -- minimizing orchestration overhead per unit of work. The comb-as-communication-medium maps to git history as signal substrate: commit frequency, size, and message precision affect how clearly workers receive information. Bee space maps to the coordination latency window -- the 6-9 second range where the hive can dispatch and receive status without state divergence.

### 2.7 Thermoregulation and Homeostasis

Brood nest temperature is maintained at 34.5-35.5 degrees C (pupae at 35.18 +/- 0.04 degrees C) through heater bees generating 209 mW/g via decoupled flight-muscle shivering, and fanner bees creating directional airflow for evaporative cooling. Winter clusters maintain a 32-37 degrees C core surrounded by a 6-12 degrees C mantle, with continuous rotation between roles. Total thermoregulation overhead: 10-15% of annual honey. No central thermostat exists; equilibrium emerges from distributed local sensing and differential individual reactions.

**Software Mapping.** Thermoregulation maps to load management. Heater bees are intensive-compute workers; fanner bees handle I/O and caching. The mantle/core structure maps to validation layers around the hot execution path. Our architecture correctly separates hot-path execution from cold-path validation, but lacks the continuous rotation mechanism -- workers do not automatically cycle between intensive and recovery states, risking burnout without the biological rest-rotation pattern.

### 2.8 Foraging Economics

Foraging cost averages 50% of energy gained. Colonies avoid nectar below 20% sugar concentration (survival filter, not preference). The marginal value theorem governs patch departure: leave when local intake rate drops below habitat average. Scout-to-forager ratio adjusts dynamically: 5% scouts during abundance, 35% during scarcity. Nectar is pre-processed in the forager's crop (concentration doubled before return), then dehydrated through a receiver-processor-storage pipeline losing 40-70% of initial water mass while retaining nearly all energy.

**Software Mapping.** Sugar concentration maps to data quality; the 20% threshold maps to minimum viable signal. The scout ratio maps to the exploration-exploitation tradeoff in task discovery. The nectar processing pipeline maps to data transformation stages with quality gatekeeping at each handoff (receiver bees rejecting dilute nectar = validators rejecting malformed input). Our system lacks explicit marginal-value departure logic -- workers complete assigned scopes regardless of diminishing returns rather than abandoning low-ROI tasks.

### 2.9 Swarm Intelligence and Emergence

Intelligence emerges from three mechanisms: stigmergy (coordination through environmental modification, not direct messaging), positive feedback (successful foragers recruit proportionally to quality), and negative feedback (pheromone decay and dance-intensity fading prevent lock-in to depleted resources). The colony's total neuron count (~900 million across 60,000 bees) rivals a rat brain, but uses distributed rather than centralized processing. Noise is functional: stochastic recruitment weighted by fitness prevents the colony from becoming locked into suboptimal strategies while maintaining concentration on good ones.

**Software Mapping.** Our pheromone broadcast implements sign-based stigmergy -- workers read ambient state rather than receiving direct messages. However, we lack sematectonic stigmergy (modification of the work artifacts themselves as communication) and temporal stigmergy (signal absence as information). The biology shows that all three levels are necessary for robust coordination. We also lack deliberate noise injection -- our system deterministically assigns work rather than allowing probabilistic task selection weighted by quality signals.

### 2.10 Colony Lifecycle

Colonies grow from founding (10,000-30,000 bees) through spring build-up (500-600 bees/day, queen laying 2,000+ eggs/day) to peak population (60,000-80,000 in June). Swarming -- colony-level reproduction -- occurs at peak, with the old queen departing with 50-70% of workers. Virgin queens emerge, mate with 12-20 drones from other colonies (ensuring genetic outbreeding), and the parent colony recovers. Supersedure (planned queen replacement) maintains continuity with overlap; emergency replacement produces lower-quality queens from suboptimal larvae. Wild colonies average 5-6 years; managed colonies 1-2 years without intervention.

**Software Mapping.** Colony founding maps to system bootstrap. Swarming maps to deliberate parallel dispatch -- the hive reproducing itself into sub-domains. Supersedure maps to planned Kairos handoff (RESUME.md protocol), while emergency replacement maps to unplanned recovery. Our handoff protocol correctly implements the supersedure model (overlap, state preservation), but we lack the swarming model -- we do not deliberately fission the system into independent sub-colonies when it reaches capacity.

### 2.11 Defense and Social Immunity

Defense operates in layers: guard bees performing stateless CHC (cuticular hydrocarbon) pattern-matching at the entrance, alarm pheromone cascades with built-in negative feedback (IPA suppresses waggle dances, preventing recruitment to dangerous locations), hygienic behavior removing 95%+ of diseased brood within 48 hours (heritable trait, 2-7 genetic loci, present in only 10% of unselected colonies), propolis as persistent antimicrobial infrastructure, undertaker bees as garbage collection (1% of population), and collective fever response raising brood temperature ~0.56 degrees C to suppress pathogen replication. Varroa destructor defeats multiple layers simultaneously through behavioral evasion, active immunosuppression, and viral vectoring.

**Software Mapping.** Guard bees map to input validation at system boundaries -- stateless, pattern-matching, no registry required. The alarm cascade with negative feedback maps to circuit breakers that both escalate alerts and suppress new work intake during emergencies. Hygienic behavior maps to quarantine-at-source: detecting and removing compromised work products before they propagate. Our system implements boundary validation and circuit breakers but lacks the propolis analogue -- persistent, ambient antimicrobial infrastructure that operates continuously without activation.

### 2.12 Multi-Modal Communication

Bee colonies operate seven simultaneous communication channels: substrate vibration (comb-borne, 12-300 Hz, detected by subgenual organ), airborne sound (hive-wide, detected by Johnston's organ tuned to 265 Hz), antennation (tactile, right-antenna preferred, 5-25 bits/sec), trophallaxis (food-based information transfer, 5-20 bits/sec per dyad with exponential fan-out), volatile pheromones (airborne chemical gradients), contact pheromones (surface-deposited markers), and visual/spatial cues. Total colony information bandwidth from dances alone: ~10^6 bits/day. The waggle dance is deliberately multimodal, broadcasting through substrate, airborne, visual, olfactory, and gustatory channels simultaneously.

**Software Mapping.** Our 7-channel pheromone system directly mirrors biological multi-channel architecture: queen status, foraging intensity, alarm level, capacity, discovery, quarantine, and hive health. The mapping is structurally correct. However, we lack the bandwidth hierarchy -- biology shows that different channels operate at different latencies (stop signals at 174 ms, queen pheromone propagation at 30+ minutes) and different reliability levels. Our channels are all updated at the same tick rate, losing the temporal differentiation that enables both fast emergency response and slow strategic signaling.

### 2.13 Genetics and Adaptation

Haplodiploidy creates 3/4 relatedness between sisters (vs. 1/2 in diploid species), explaining worker altruism through Hamilton's rule. The complementary sex determination (CSD) locus enforces genetic diversity: homozygous offspring are inviable, making inbreeding lethal at the developmental level (Beye et al., 2003). Queens mate with 12-20 drones, creating patriline diversity that produces 36% more waggle dances, 62% more waggle runs, and 2/3 higher scout return rates compared to single-patriline colonies (Mattila & Seeley, 2007). Hyperpolyandrous colonies show 36% lower Varroa infestation. Epigenetic mechanisms (DNA methylation, histone acetylation) determine queen vs. worker fate from identical genomes based on diet within a 3-4 day critical window.

**Software Mapping.** Polyandry maps to deliberate worker diversity -- seeding the worker pool from multiple model lineages, training regimens, and code versions. Our system currently lacks anti-monoculture enforcement. The CSD mechanism suggests we should make homogeneity actively harmful: if all workers converge to identical strategies, the system should degrade rather than simply losing resilience passively. Epigenetics maps to worker role plasticity -- the same worker code deployed in different configurations based on pheromone signals, without restart.

### 2.14 Navigation and Memory

Bees employ a time-compensated sun compass with magnetic backup, path integration (dead reckoning), and landmark-based route memory, with evidence for map-like spatial representations emerging from integration of multiple routes. Mushroom bodies (corpora pedunculata) serve as the learning center, with Kenyon cells forming microglomeruli whose synaptic density increases with long-term memory formation. Working memory lasts ~5 seconds; counting capacity is limited to 4 items; bees can learn addition and subtraction using abstract color-coded rules. Time-place learning enables scheduling foraging to match nectar secretion peaks across up to 9 daily time slots. Neonicotinoid exposure impairs olfactory learning by ~50% while leaving color learning intact (modality-specific vulnerability).

**Software Mapping.** Navigation maps to worker task-routing: the hierarchical system of compass (global objective), path integration (local progress tracking), and landmark calibration (checkpoint verification) is a template for how workers should navigate complex, multi-step scopes. The time-place learning capacity suggests workers should learn temporal patterns in task availability and pre-position for predicted workloads. The modality-specific pesticide vulnerability warns that system degradation may affect some capabilities while leaving others intact -- partial failure, not total collapse.

### 2.15 The Beekeeper-Colony Relationship

The beekeeper provides shelter, disease monitoring, and genetic refreshment while the colony retains operational autonomy. Over-management creates dependency: routine chemical treatment without monitoring breeds resistant pests and preserves weak bees. The one-third rule governs harvest: take surplus, leave two-thirds for colony survival. Inspection frequency follows a spectrum from weekly (new colonies) to yearly (top-bar hives). Non-invasive monitoring (entrance reading, hive weight, sound analysis) reduces inspection cost. Wild colonies with genetic diversity show 80% annual survival; managed colonies without intervention show 50%.

**Software Mapping.** The beekeeper is the user -- the only authority from outside the hive. Kairos is the swarm steward, not the hive mind. The Apis loop is continuous light inspection. The one-third rule maps to measured value extraction: the system should produce surplus beyond what it needs to maintain itself. The treatment-dependency trap maps to over-engineering: excessive intervention (too-frequent deploys, too many guardrails, too much monitoring) creates systems that cannot self-correct. The biology's intervention framework -- act only when collapse is imminent, observe when adaptation is in progress, trust self-correction for normal fluctuations -- is directly applicable to system operations.

---

## 3. Architectural Validation

### 3.1 Where We Follow Biology Correctly

**Broadcast-not-query pheromone.** Apis writes the 7-channel pheromone every tick; workers read it at scope start. This correctly implements the QMP distribution model -- continuous broadcast, ambient sensing, no request-response overhead. (Implementation: `docs/APIS.md`, pheromone architecture.)

**Anonymous, interchangeable workers.** Workers are generic processes with no persistent identity. Work is the artifact, not the worker. This matches the biological reality that individual bees are expendable; the colony survives worker loss. (Implementation: `docs/WORKER-TEMPLATE.md`.)

**Orchestrator as queen, not as manager.** Kairos produces strategic decisions and allocation signals but does not micromanage worker execution. The queen does not assign tasks; she broadcasts state. (Implementation: `CLAUDE.md` standing orders, `docs/ORCHESTRATED-HIVE.md`.)

**Multi-channel information architecture.** The 7-channel pheromone directly maps to the seven biological communication channels: status, foraging intensity, alarm, capacity, discovery, quarantine, and health. (Implementation: `docs/HIVE-ARCHITECTURE.md`.)

**Hexagonal scope design.** Scopes minimize orchestration boundary relative to work content, matching the hexagonal wax-efficiency principle. (Implementation: `docs/ALLOCATION-PROTOCOL.md`.)

**Handoff as supersedure.** The RESUME.md protocol implements planned succession with overlap -- the current Kairos prepares state for the next instance. (Implementation: `RESUME.md`.)

### 3.2 Where We Diverge

**No pheromone decay.** Biology relies on exponential QMP decay (36-61 minute half-life) to detect queen loss. Our pheromone persists indefinitely once written. Workers cannot distinguish "Apis wrote this 5 seconds ago" from "Apis wrote this 3 hours ago and has been dead since." This is the single most critical divergence.

**No cross-inhibition.** Competing approaches to a problem run in parallel without mutual suppression. The biology shows this causes deadlock under equal-quality competition. We need stop signals.

**No response-threshold self-selection.** Workers are explicitly assigned rather than self-selecting based on stimulus sensitivity. This eliminates the emergent specialization and crisis flexibility that threshold-based allocation provides.

**No anti-monoculture enforcement.** All workers can be identical without system penalty. The CSD mechanism shows that homogeneity should be actively harmful, not merely suboptimal.

**No tremble-dance backpressure.** When processing capacity is saturated, workers fail silently rather than broadcasting a capacity signal that suppresses further recruitment. The system over-commits rather than self-regulating.

**No deliberate noise.** Task assignment is deterministic. Biology shows that stochastic recruitment weighted by quality prevents lock-in and maintains exploration of alternatives.

---

## 4. Gaps and Recommendations

Ranked by impact on system resilience:

1. **Implement pheromone decay** (Critical). Add timestamps to pheromone channels. Workers should treat stale pheromone (>N ticks old) as equivalent to absent queen signal and enter safe-mode behavior. This is the colony's primary failure-detection mechanism and we lack it entirely.

2. **Add cross-inhibitory stop signals** (High). When a worker detects that an approach is failing, it should emit a signal that suppresses recruitment to that approach. Without this, the system cannot reach consensus when multiple competing strategies exist.

3. **Implement backpressure signaling** (High). Workers encountering processing delays should broadcast a tremble-dance analogue that suppresses further task generation. This prevents the over-commitment cascade that currently manifests as queue overflow.

4. **Introduce response-threshold self-selection** (High). Workers should self-select tasks based on configured sensitivity profiles rather than receiving explicit assignments. This produces emergent specialization while preserving crisis flexibility.

5. **Enforce worker diversity** (Medium). Track worker lineage diversity and degrade system health metrics when homogeneity exceeds a threshold. The CSD mechanism shows that monoculture should be structurally penalized.

6. **Add temporal channel differentiation** (Medium). Different pheromone channels should operate at different update rates: alarm channels at sub-second latency, strategic channels at minute-scale. This matches the biological bandwidth hierarchy.

7. **Implement marginal-value departure** (Medium). Workers should abandon scopes whose return rate drops below the system average rather than completing them regardless of diminishing value.

8. **Add propolis-equivalent persistent defense** (Low). Continuous, ambient input validation that operates without explicit activation -- infrastructure as immune response, not reactive incident handling.

9. **Implement worker rest-rotation** (Low). Workers should cycle between intensive execution and recovery states, matching the mantle/core rotation in winter clusters. This prevents burnout and maintains sustained throughput.

---

## 5. Conclusion

The bee analogy holds. Across 15 domains of colony biology, every mechanism we surveyed has a direct structural parallel in distributed software systems. The parallels are not metaphorical -- the mathematical structures of quorum sensing, cross-inhibited competing accumulators, threshold-based self-selection, and stigmergic coordination are identical in biology and computation. Both solve the same fundamental problem: coordinating large numbers of limited agents to achieve collective intelligence without central authority.

Where our architecture follows the biology -- broadcast pheromone, anonymous workers, orchestrator-as-queen, multi-channel communication -- the system exhibits the robustness that 100 million years of evolution has refined. Where we diverge -- no signal decay, no cross-inhibition, no anti-monoculture enforcement, no backpressure -- the divergences are not deliberate design choices but architectural debt.

The recommendation is unambiguous: where the implementation diverges from the biology, we should change our architecture, not the analogy. The colony has been debugged by natural selection across geological time. Our system has been debugged for months. The burden of proof falls on the divergence.

---

## References

Beshers, S. N., & Fewell, J. H. (2001). Models of division of labor in social insects. *Annual Review of Entomology*, 46, 413-440.

Beye, M., et al. (2003). The gene csd is the primary signal for sexual development in the honeybee. *Cell*, 114(4), 419-429.

Dong, S., et al. (2023). Social signal learning of the waggle dance in honey bees. *Science*, 379(6636), 1015-1018.

Hales, T. C. (1999). The honeycomb conjecture. *Discrete and Computational Geometry*, 25(1), 1-22.

Keeling, C. I., et al. (2003). New components of the honey bee queen retinue pheromone. *Proceedings of the National Academy of Sciences*, 100(8), 4486-4491.

Mattila, H. R., & Seeley, T. D. (2007). Genetic diversity in honey bee colonies enhances productivity and fitness. *Science*, 317(5836), 362-364.

Nowak, M. A., & Wilson, E. O. (2010). The evolution of eusociality. *Nature*, 466, 1057-1062.

Peso, M., et al. (2016). Queen mandibular pheromone composition affects worker retinue response. *PLoS ONE*, 10(7), e0156027.

Robinson, G. E. (2002). Genomics and integrative analyses of division of labor in honeybee colonies. *American Naturalist*, 160(S6), S160-S172.

Seeley, T. D. (2010). *Honeybee Democracy*. Princeton University Press.

Seeley, T. D., et al. (2012). Stop signals provide cross inhibition in collective decision-making by honeybee swarms. *Science*, 335(6064), 108-111.

Tanner, D. A., & Visscher, P. K. (2006). Do honey bees tune error in their dances in nectar-foraging and house-hunting? *Behavioral Ecology and Sociobiology*, 59(4), 571-576.

*Additional sources: 200+ peer-reviewed papers cited across the 15 domain research files in `.swarm/beliefs/bee-research/`.*