Gene Hive

What happens when the same inheritance shapes both the creature that carries the food and the decisions that send it out to find it?

A worker may be quick but expensive to sustain, well protected but heavy, or especially responsive to a hungry larva nearby. Its own needs compete with foraging, care, exploration and rest, and those choices decide how much reaches the stores, how quickly brood develops and whether the nest has room to grow.

The interesting scale is the space between one insect and the whole colony: a small difference in response can change who does the work, while a shortage or a predator can change which traits are useful.

• Behaviour · 17 genes Priorities, persistence, risk tolerance, cooperation, kin preference, recruitment, learning rate and willingness to answer colony demand.
• Morphology · 9 genes Eyes, legs, storage tissue, mouthparts, antennae, hydration and immune tissue, carapace and defensive spines.
• A finite tissue budget No insect gets every organ at full size for free: oversized allocations are scaled back, and body mass and metabolism make every body a trade-off.
• A painted part rig Bodies are assembled from artist-editable pieces, so expressed traits are visible, with age and injury changing the look further.

Eyes and antennae affect what a worker can sense, legs how it moves, storage what it carries, mouthparts how it handles food, and armour and spines how it fights. The inspector shows inherited and expressed traits side by side, so a large genetic allocation and the body that actually develops can be told apart.

The neural network has 11 inputs, four hidden neurons and 11 outputs. It reads signals such as hunger, thirst, fatigue, threat, resource scarcity and nearby distress, and adds a bounded adjustment to authored utility scores.

• It shapes intent The network nudges which goal wins. Feasibility checks, pathfinding, resource accounting, job claims and action execution stay deterministic rules.
• Safety stays authored A separate safety system steps in when an urgent need or a threat calls for recovery or escape.
• Learning is not inheritance The output layer learns during an insect’s life, but offspring inherit the stored genotype, not the parent’s learned weights. Evolution across generations and adaptation within one life stay separate.

The diagnostics keep that boundary visible: they separate policy decisions from safety, defence, hauling and other authored interventions, and report how often the neural adjustment actually changed the chosen goal.

The surface is both the source of resources and the source of pressure.

• Nectar, water and resin Collected, carried and stored explicitly. Carrying costs something, and resources have to physically reach the colony.
• Pheromone trails Local cues for home, food and water that evaporate over time, so no insect ever knows the whole map.
• Weather, seasons and daylight They change the environment and its demands. Pond water is finite, recharged by rain and lost to evaporation.
• Plants and soil Regrowth, biomass, decay and fertility are linked. Dead insects keep their identity, body mass and cargo until recovered or decomposed.
• Fungus and injury Costs beyond hunger and thirst. Rest and mutual care help, but contact can also spread infection.
• Spider predators They stalk, attack, recover and retreat. A defeated spider becomes food, but hauling and processing it takes work.

Four pressure presets run the colony under different conditions:

• Calm Observation
• Seasonal Challenge
• Predator Focus
• Disease Focus

Below the surface, the colony lives on its own hex layer, connected through entrances. The Nest view shows nurseries, food and water stores, material storage, resting space and the other chambers, and these spaces matter: brood needs nursery capacity and care, stored resources need room, and digging competes with every other job.

Reproduction draws on healthy, eligible adults and the colony’s resources. Contribution credit influences who becomes a parent, and brood has to be funded and raised, so every birth is part of the colony economy. Crossover and mutation create new combinations of behaviour, body and brain, and occasional immigrants add variation when population or genetic diversity runs low.

Social behaviour combines inherited tendencies with authored mechanisms: food sharing, kin-biased helping, grooming and alarm recruitment give individual differences a way to matter. Whether a worker responds to care requests depends on what it perceives nearby and on its inherited threshold, and increasingly urgent requests recruit less responsive workers too.

A changing gene distribution is not, by itself, proof of adaptation.

That is why the project measures whether traits are expressed, whether they affect work and survival, and whether the insects carrying them actually contribute to later generations.

The interface is part of the experiment: it should explain what happened.

• Specimen inspection A worker’s body, needs, current goal, genome and neural activity in one place.
• Evolution statistics Population, generations, genetic variety, trait distributions, ancestry and likely breeding candidates.
• Decision diagnostics Where decisions came from, how much the network influenced them and how time was spent across goals.
• Environment statistics Climate effects, pond conditions, fungal load and the cost of predation.
• Resource statistics Incoming supplies separated from upkeep, brood and construction, next to physical storage capacity.
• Overlays and history Map overlays and an event log for exploration, traffic, resources, danger and change over time.

Pause, speed controls, seed changes, scenario resets and save/load support watching at different timescales, and interface scaling and reduced motion keep a dense simulation readable.

The simulation advances in fixed 60 Hz steps with seeded random streams and explicit state; fast-forward changes how many steps run, not their size. The authoritative colony model is plain C# and runs without a scene tree. Godot handles presentation, input and the camera, and the views only read simulation state: opening a statistics panel or switching an overlay must never change the colony’s outcome.

An offline genetic algorithm searches for policies with selection, crossover and mutation; optional NSGA-II selection compares several objectives instead of reducing every trade-off to one ranking. Broad searches can use a cheaper proxy, but promising policies also face the real headless colony runtime, because doing well in a simplified evaluator is no evidence of doing well in the game.

Seeded comparisons, separate validation banks and frozen-versus-learning runs separate inherited performance from lifetime adaptation, and test-only tools hunt for policies that collect reproductive credit without doing the useful work.

Does a policy help the colony under the rules it will actually meet, and does that still hold when the scenario changes?

Gene Hive is also an experiment in human-directed, AI-assisted development. Coding agents help with implementation, investigation, testing and documentation; the running simulation itself is explicit C# rules, genetic operators and a compact neural network.

Gene Hive is a playable foundation in active development. The next questions are about behaviour and selection:

• Cooperation When does helping others repay its cost?
• The brain How much useful control does the network actually exert?
• Pressure Do weather, disease and predators produce meaningful differences across generations?

The current tools make these questions inspectable: trait distributions reveal differences a colony average hides, decision traces show when the policy had a real choice, and resource and predation statistics show what that choice cost. The direction is a colony whose changes can be followed from the body of one worker to the history of the nest, backed by evidence whenever the simulation seems to have learned something.