Thegenome

Chosen after surveying Avida, Aevol, NEAT and Banzhaf's ARN: take the tractable encoding, drop their selection machinery, keep the biological growth mechanism.

• An explicit network. The genome is a list of nodes plus directed weighted edges — like NEAT, but with no innovation numbers (those exist for crossover, and cells divide asexually).
• Fixed interface, evolvable interior. Every cell carries all of §3's sense-inputs and effector-outputs as nodes; evolution wires the hidden network between them. An unused effector simply evolves near-zero drive — nothing presets a guild.
• The world selects. NEAT and Aevol both assume a fitness target; we take only their representation. Survival and reproduction in the ecology are the only selection.

Why duplication and not NEAT's connection-splitting: real regulatory networks grow by duplicating a regulator then diverging it, and that mechanism alone produces lifelike (scale-free, modular) topology — structure for free, before any selection. It is load-bearing, not flavour.

Continuous and recurrent, with memory built in.

• One synchronous update per tick. Each node's new value = a bounded function of its weighted inputs plus the previous tick's hidden state. Nothing is run to equilibrium — state persists.
• That persistence is the memory (H7). A small cell can't read a spatial gradient (§3), so it compares now against its remembered past and run-and-tumbles; diel rhythm can also be learned. The recurrence is what makes that possible.
• Depth costs ticks. A signal crossing a deep chain takes several ticks to arrive — biologically reasonable latency, and it keeps each tick's compute O(nodes).

Mutation operators, applied at division from the dedicated mutation RNG stream (§4).

• Duplication-and-divergence — the complexity engine. The paralog starts neutral; divergence comes from ordinary point mutation afterward.
• Point — perturb edge weights, node biases, and the marker tag (Gaussian).
• Add / remove edge, and node deletion — so networks can shrink, not only grow. Aevol's lesson: deletions are what bound bloat.

• Copy then mutate. At division the genome is duplicated and the copy mutated. Daughters share a genotype.
• Differentiation without genetic difference. The recurrent state is inherited and may be split asymmetrically (§1), so two cells with the same genome can behave differently — a colony can specialise its cells by state and position alone.
• The marker tag is 3 evolvable genome scalars (the emitted barcode); emission rate is an output; the response is wiring from the 3 marker-sense inputs. Kin recognition, greenbeard cooperation, and cheating all fall out of co-encoding tag and response.

How outputs become behaviour, and where a run starts.

• Outputs are target drives. Effector outputs set target efforts and allocations; §3.3's standing machinery relaxes toward those targets and pays maintenance at its current level — so machinery is slow and hysteretic, and idle capacity still costs.
• Minimal seed. A run starts from sparse, near input→output genomes with small random weights — no hidden structure handed in, so behaviour and guilds emerge.