Tecto, a Digital Fish from the Silicon Cambrian, A 128-Neuron Recurrent Network Recapitulates the Optic Tectum

A GRU gated recurrent network with 128 hidden units, trained atop a frozen, self-supervised visual encoder, achieves simultaneous foraging and obstacle avoidance in three dimensions using only binocular vision -- without semantic labels, depth maps, or any coordinate information. We call it Tecto, a 128-neuron digital fish from the Silicon Cambrian. Within its training distribution, it reaches a 99.7% food-capture rate with a 1.3% obstacle-collision rate, and tracks randomly-walking food with an 82.8% success rate despite never being trained on moving targets. Systematic circuit analysis reveals that the network spontaneously converges to the same functional organization as the vertebrate optic tectum: the update gate z encodes behavioral intensity (neuromodulatory gain control), the reset gate r is structurally silent, foraging and avoidance are encoded along nearly orthogonal hidden-state directions, and the output matrices implement a distributed population-vector decoder. Ablating the top-16 hidden dimensions drops the food-capture success rate from 98.0% to 80.7%; ablating the bottom-16 has zero effect, confirming approximately 12.5% of parameters as silent synapses. Fifteen generations of evolution fail to improve the pretrained brain -- in a 594,468-dimensional parameter space the uphill probability of random mutation is below 10^{-5000}. Reverse validation through targeted lesioning provides independent evidence for the model's biological fidelity. We systematically damaged ten circuit loci spanning sensory, gain-control, recurrent, and motor-output pathways. Each lesion produced a behavioral phenotype matching the clinical presentation of the corresponding biological damage: b_z shift generates a continuous bradykinesia-to-hyperkinesia spectrum with preserved task competence; unilateral W2 row ablation recapitulates contralateral saccadic deficits; sensory-column perturbation at delta/theta frequencies produces direction-specific resonance; and coherent noise in only 5% of neurons is sufficient to collapse population-code averaging. Every lesion maps to a known computational function in the characterized circuit. We propose that a set of physical constraints -- binocular vision, continuous time, finite neuron budget, real-time motor output -- defines a low-dimensional feasible solution space whose stable solution is the functional organization of the optic tectum. This hypothesis makes a specific, falsifiable prediction: if the same constraints are imposed on a different encoder (e.g., CLIP), a different recurrent architecture (e.g., LSTM), or different motor dynamics, the same structural convergence should re-emerge. Should it fail to do so, the hypothesis is falsified. These substitution experiments have not yet been performed; the hypothesis currently rests on a single data point.