Six models wake up at 06:00. One swarm stays awake for 18 hours. Zero humans touch a thing. Welcome to Batonical — a closed-loop control system where spiking transformers drink from a hyperspherical manifold, accumulate evidence on leaky integrate-and-fire membranes, and fire precise hardware commands when biology demands it.
Six checkpoints load in parallel. Each runs a smoke-test forward pass to prove its graph is alive.
Multispectral imaging, soil chemistry, RGB vision. Each modality gets its own encoder.
scientific_collection_manifold.ptorganic_manifold_weights_v1.pthorganic_distilled_automator_scripted.ptEach encoder L2-normalizes its output. The three vectors land on a unit hypersphere. Drag to rotate.
The Manifold Archive looks up the top-3 most similar past biological states and pulls them into context.
The LIF membrane accumulates evidence. Six action channels, each leaking at rate 0.15. Threshold: 0.85. No spike yet — the plant is thinking.
Leaky Integrate-and-Fire: membrane = membrane × (1 − leak) + current, fire when membrane ≥ 0.85, reset on fire.
Threshold crossed. The Semantic Router dispatches the spike to a real hardware driver.
| action | driver | bus |
|---|
Same telemetry in. Same action out. The 7 MB student matches the 1.5B-parameter Qwen teacher — and runs without the transformers library.
Three TorchScript-compiled models loaded straight from .pt files. The class definitions never left the training notebook.
18 hours. 24 ticks. The plant is healthy. The LTM is full. Tomorrow the leak rate resets, but the archive does not.
— End of mission. The greenhouse sleeps. The swarm doesn't.