NeuroTerrarium

The embodied terrarium the user first sees is a 48-neuron synthetic fixture with hardcoded weights and no connectome in it at all. What keeps this project on the scale is the separate Phase 3 instrument: a real MaleCNS 51-body giant-fibre input microcircuit, bundled with its source queries and checksums, converted to signed weights from measured contact counts and stepped through the same leak/threshold/refractory LIF for a bounded pulse experiment against matched controls and ablations. That is L1 - real wiring, real membrane dynamics. Not L2: the only fitted parameters anywhere are 49 linear readout coefficients trained by gradient descent in learning.ts, on the synthetic fixture rather than the connectome, and the app's own mode table says perturbations never edit recurrent edges. The project also publishes its own negative result - the biological microcircuit and all five rewired controls respond identically to the default pulse - and that honesty should carry into the index rather than count against it.