Boundary surfaces stay explicit
Projection-style I/O keeps the public interface readable and comparable even while the hidden computation changes aggressively.
Architecture line
INSTNCT treats topology as part of the learnable object.
INSTNCT is Vraxion's current architecture line: a gradient-free self-wiring system that changes its own directed graph instead of learning inside a fixed layer stack.
The public architecture claim is intentionally narrow: explicit I/O boundaries, a sparse signed recurrent graph, persistent state across ticks, and mutation-selection training.
How the line is built
Fixed boundary, learnable substrate, persistent runtime.
The line is best understood as an architecture essay, not a list of knobs.
The hidden graph is the main search object
Evolution rewires sparse routing and compact neuron roles instead of only nudging dense internal weights.
Runtime state persists across ticks
Charge and state survive long enough for computation to unfold across a recurrent substrate rather than collapsing into one pass.
Public status stays split
Shipped defaults, validated alternatives, and still-experimental branches do not get mixed together in the public story.
Main architectural fork
One shipped line, one serious non-default alternative.
The key comparison is not a zoo of equal options. It is the default line against the strongest validated alternative.
Shipped current mainline
Projection-style boundary
input -> input_projection -> hidden signed graph -> output_projection -> output
persistent charge and state across ticks
Strongest validated alternative
Tentacle I/O
input -> first V neurons -> hidden graph -> last V neurons -> output
direct boundary inject terminal charge as logits
| Aspect | Shipped line | Validated alternative |
|---|---|---|
| Boundary handling | Fixed projection-style input and output interfaces. | Boundary neurons act directly as the I/O surface. |
| What stays fixed | Boundary interface shape remains explicit. | Boundary placement stays fixed, but routing becomes more direct. |
| What is learnable | Hidden graph plus compact neuron controls. | Hidden graph plus routing behavior closer to the boundary. |
| Current status | Canonical line. | Strongest validated alternative, not the default. |
Fixed versus learnable
The architecture rule stays simple: keep the public boundary explicit and let the recurrent substrate carry the learnable burden.
- Input and output interfaces stay explicit.
- The hidden graph and compact neuron controls remain learnable.
- Charge and state are runtime dynamics, not just stored parameters.
Theory capsule
Inference emerges as the fixed point of destructive interference.
Signal enters a recurrent substrate, incompatible paths suppress one another, and the surviving residue is read out as inference. The theory explains the line, but it does not replace architecture status or proof.