VRAXION is building a gradient-free self-wiring architecture called INSTNCT. The learnable object is the graph itself, not a fixed layer stack, and the public story stays organized around one contract: what is shipped on main, what is validated but not promoted yet, and what is still experimental.
v4.2.0 stable
v5.0.0 Public Beta
instnct/ is the repo path, not the release label.
VRAXION is a company and engineering program centered on one active architecture line: passive I/O, a signed self-wiring hidden graph, persistent internal state, and mutation-selection training.
The only settings that count as live defaults are the ones actually shipped in instnct/model/graph.py. That mainline currently includes per-neuron theta/decay and nonnegative charge dynamics.
The strongest schedule result so far is voltage medium leak at 22.11% peak and 21.46% plateau. The strongest compact learnable controller so far is the 3-angle decision-tree schedule at 20.05%.
The main live target is context-dependent task learning: windowed input injection, associative memory, and stronger evaluation for nontrivial tasks.
main.These are the quick checks that keep the public story aligned with the code.
python -m compileall instnct tools python instnct/tests/test_model.py python tools/check_public_surface.py
The public stack is intentionally shallow: use VRAXION Home for orientation, then jump to architecture, evidence, engineering, or timeline by question.
