What happened.
What remains open.

Growth becomes physical habitat.

Three related, supplied colony genomes ran for 180 model seconds with native chemistry and MuJoCo geometry. They built over 1,600 parts, mostly roots beneath the supplied terrain, with branches and leaf surfaces above or alongside it. A restored checkpoint reproduced the subsequent physical, chemical and developmental state exactly in the same runtime.

Live tissue converted to detritus while the scaffold persisted. An experimentally removed leaf transferred its remaining material into another lineage’s gut, where the common reaction system digested it. End-to-end elemental residuals were below 10⁻¹² in the model’s abstract units; the energy residual was also below 10⁻¹².

This establishes joined mechanisms. The founders and their endowments were engineered. The transfer was supplied by the experiment, and the run contained no neural controller. Autonomous predation, reproduction and ecological selection were not demonstrated.

Changing terrain changes transport.

Static additions and removals now change the chemical field’s geometry. A newly built solid conservatively displaces covered material; reopening a passage restores fluid space. A physical wall experiment retained total mass exactly, blocked cross-wall transport, then resumed transport after removal.

The field uses a finite grid. Tiny leaves and cracks can fall below its resolution. Displacement is a numerical material-conservation rule, not a fluid pressure or momentum simulation.

Inheritance specifies capacities and construction.

A small genotype compiles into actual enzyme expression, growth dimensions and neural port gains. Offspring carry source and ancestry hashes. Learned memories, current neural activity and adult physiology remain outside the genome.

Mobile morphology can also vary across residents in the same world: actual trunk and limb dimensions, antennae, mass, friction and bounded motor parameters. The current mobile interface remains six-legged. These are mechanisms for diversification, not evidence that useful specializations have evolved.

A measured scaffold with explicit assumptions.

Apple Metal and AMD GPU kernels advance the large graph. Per-resident activity, adaptation and support state are private. The graph’s anatomical provenance does not make the model’s physiology or behavior a biological reconstruction.

Personal history can inform imagined futures.

A recurrent predictor trained on 184,320 aligned transitions from actual full-graph developmental runs. Whole worlds were held out. The trained model beat persistence on feature and physiology prediction at the reported horizons; a Rust implementation reproduces its forecasts closely and maintains private recurrent state.

Changing the supplied action history changed forecast quality. The strongest discrimination was in neural features and angular motion. Energy, gut contents and support showed much weaker action sensitivity.

Private counterfactual queries now execute. Useful planning and improved long-term regulation remain unestablished. A plausible forecast is not sufficient evidence to trust a policy built around it.

More neurons have not solved foraging.

Short developmental probes have shown effects of neural input and learning. Longer resident observations still expose reserve depletion and ineffective activity. Sustained autonomous feeding, useful manipulation and durable individual strategies remain unmet capabilities.

Biological wiring also has to earn its role: a matched rewired graph performed better on ingestion in one earlier controlled comparison. That makes the controls informative and the architecture revisable.

A circuit can acquire a traceable descendant.

A blueprint duplicated a 3,627-neuron learning-circuit template and its declared incoming and outgoing connections. The resulting graph has 168,749 neurons and 26,557,080 directed edges. All original neuron metadata and ancestral-to-ancestral edges remained exact.

The derived graph executed on AMD hardware, and its duplicate cells responded to sensory input. Added outputs also changed incoming normalization on 3,152 ancestral targets, which the artifact records explicitly.

This establishes executable structural variation. The copy is a synthetic intervention; useful specialization, developmental biology and behavioral improvement remain unestablished.

From company to interdependence.

Shared physics, sensory signals and persistent structures provide places to begin. Embodied reproduction, dependent young, resource-funded nursery care and evolved divisions of labor are still to be built.

We plan to compare individual and group outcomes, hold resource budgets constant, intervene on communication and construction cues, and distinguish learned task preferences from reproductive specialization.

A mouthful becomes private physiology.

Fresh worlds can now bind mobile body and gut compartments to the same native chemistry as growing colonies and finite movable stores. ATP funds active forces and signals. Actual contact near the mouth can transfer material into the gut; digestion and absorption then affect the resident’s internal state.

A physical mouth-contact experiment ingested 0.008 abstract material units and absorbed 0.005613 reserve units. Matched side-contact and no-eating controls transferred nothing. All three scenarios restored and continued exactly.

Shared-store requests receive proportional shares under scarcity. Material depletion changes object size, mass and inertia at declared boundaries. A second circuit generation also now executes, with 172,376 neurons and ancestry preserved through both births.

The feeding experiment supplied actions and founding positions. Learned foraging under this physiology remains under investigation. Actuator activation uses an engineered energy law; mechanical work is measured separately. Nutrient recycling, reproduction and complete chemical-to-physical mass coupling remain unfinished.

One resident’s material can reach a root.

Mobile outlets now leave finite physical deposits from private gut and body chemistry. Constructed roots acquire mineral and carbon through actual contact with these objects. Empty deposit slots can be reused at new positions; full capacity leaves material in the resident.

Three residents generated deposits under supplied quiet actions. A declared positional intervention brought one deposit against a root, which acquired 0.000455 mineral and 0.000682 inorganic-carbon units after one physical tick. A matched deposit 0.8 metres away transferred nothing. Both branches continued exactly after restore, with chemical accounting residuals below 10⁻¹⁰.

This joins spatial material return and acquisition. The residents did not learn to deliver nutrients. Microbial decomposition, soil transport, ecological stability and complete mass–inertia coupling remain open.

Absorption was not food acquisition.

Matched AMD developmental runs advanced 48 residents through the new shared chemistry, using either the canonical graph or its second structural generation. Each completed 115,200 resident training steps. Neither acquired a finite food packet; ingestion-disabled evaluations consequently matched.

The recorded nutrition came from digestion of material already in the initial guts. Neural-feature interventions changed movement, but this experiment established neither autonomous feeding nor an advantage for the measured wiring.

The next nursery starts with empty guts and varied physical resource encounters. Its learning signal must distinguish an ingestion request, a real mouth contact, acquired material and later absorption.

Possible futures enter the motor loop.

A recurrent model trained on 184,320 full-connectome developmental transitions now supplies bounded action evidence in a fresh recycling habitat. Each resident keeps its own experienced state. Imagined action continuations are separate queries and do not become personal experiences.

The artifact pins its sensory normalization and quarter-second observation interval. Its action discrimination is stronger for neural features and angular movement than for energy and gut contents. Better planning or sustained regulation has not been demonstrated.

Growth belongs in the checkpoint too.

A separate research world developed physical parts, returned material, and accumulated private motor learning and predictive memories. After 70 ticks it saved, advanced 12 more, restored, and repeated that continuation. Every recorded state section and the native neural checksum matched exactly.

The check exposed a subtle error: sorting JSON keys changed the order of floating-point tissue reductions. New snapshots preserve that order. It also caught holds that still referenced consumed objects. Both mechanisms are fixed.

This finite replay does not establish uninterrupted operation. A live snapshot timeout required recovery from an earlier complete checkpoint; the missing interval is explicitly recorded. The transport now queries committed request receipts without repeating neural work, and interrupted saves retain local private state for recovery.

The body was predictable. The goal fit failed.

Native GAM models used 276,480 transitions for training, 92,160 from held-out worlds for validation, and a complete 122,880-transition episode for final holdout. Effort prediction strongly beat its mean baseline. The body-law residual model only narrowly beat the baseline on the whole held-out episode.

The native goal-progress optimization did not certify a stationary fit, so no goal model was minted. The behavioral audit is equally concrete: 90 of 96 resident-episode trajectories reached mouth-material contact, yet every episode remained energy-negative.