ONE BRAIN · FIXED UNIT-STRENGTH WIRES · LEARNING BY RECONNECTION

Ternary Triple-Pendulum · Structural Learning Lab

A three-link inverted pendulum under the full coupled point-mass equations. A single undifferentiated recurrent brain of ternary neurons learns to balance by rewiring addresses. Each neuron receives exactly three distinct unit-strength inputs, emits −1, 0, or +1, and has no trainable weights, gains, or biases. Fixed +1, −1, and 0 tonic signals are available. Nine readout neurons are grouped into three fixed motor sums.

Seed Neurons Trials/s Start tilt ±° Motor max Training cases Case rotation

Press Start learning. The initial motor neurons are silent; the network must discover stabilizing connections.

Rewire trials0

Accepted changes0

Training mean—

Unseen mean—

Unseen worst—

Unseen full 5 s—

Physical triple pendulum

Current rewired brainModel upright axis

Live preview repeats independent initial conditions. Freeze & test to stop all rewiring and inspect one uninterrupted episode.

No evaluations yet. Tests do not change the brain.

Connectome · actual source addresses

Blue wires start at a fixed ternary sensor/tonic; purple wires are recurrent. Every wire has strength +1. Green-rimmed cells N0–N8 feed the motors in groups of three. Gray cells are available as recurrent intermediates.

Neuronal rule

Nᵢ(t+1) = sign(Sₐ(t) + Sᵦ(t) + S𝒸(t))

Each source is one of:
  • quantized angle / angular velocity signals
  • fixed tonic +1, −1, or 0
  • another neuron (possibly recurrent)

τⱼ = torqueMax × (N₃ⱼ + N₃ⱼ₊₁ + N₃ⱼ₊₂) / 3

Learning changes only the 3 source indices.

Frozen-brain stress battery NO LEARNING DURING TESTS

Every condition uses the same fresh, seeded starting cases. The wiring is copied and frozen for the entire battery. The baseline tests whether it balances; the ablations test why. Changing sensors in a test never changes the stored brain.

Cases per test Include PD reference

Not run yet. Press Run full battery after the brain has learned, then paste its BX report back into chat.

Condition	Mean	Worst	Full survival	RMS tilt	Avg \|motor\|
No test results yet.

Each row reports duration relative to its own test cap (5 s or 15 s). RMS tilt is the pooled angular deviation from vertical, in degrees; motor effort is the mean absolute applied generalized torque across three axes. Compare ablations with the normal 5-second baseline, not with the 15-second run. Long survival alone is not proof of active feedback: pay particular attention to all sensors removed and motors off.

Learning curve · unseen tests

Validation meanValidation worstTraining mean

Twenty-four fixed validation initial conditions are never used to accept a rewire. A separate 30-case fresh test is generated on demand.

Experiment log and biopsy

Ready. Structural trial-and-error search, single brain, no population.

Interpret carefully: this is reward-guided trial-and-error structural optimization within one ongoing brain, not backpropagation and not yet a biologically local synaptic-plasticity rule. Fixed tonic input signals are mathematical offsets (although the neurons have no learned biases). Sensor thresholds, readout groupings and motor torque are fixed by the experimenter. The physical model is a coupled three-link pendulum with massless rods and point masses; motors apply generalized torques to the absolute link coordinates.