GENESIS OPERATOR-FORMING UNIT · SERIAL ···· · LAB

WATCH IT FORM A MODEL.▊

Here is the simple experiment: hide 97% of a moving 3D body and ask Genesis to keep a complete internal model of it. If the blue prediction stays aligned with the real body, it is inferring what it cannot see—not merely copying the input. The other tabs expose the underlying operator-forming experiments.

The challenge Genesis receives 983 scattered measurements. It must infer the other 31,785 voxels while the body moves.

not a recording○ WAITING FOR A LIVE FRAMEASUS GPU receipt appears after START

last origin compute—fresh server UTC appears here

one-use response nonce—changes on every computed frame · API is no-store

Verify it yourself: press START, then SURPRISE IT. Your browser creates a nonce only when you click; the ASUS server derives an unseen movement from it and the live prediction must recover.

1A BODY GROWSIt began as one active cell. Prediction pressure grew it into an 802-cell 3D organism.

2WE BLINDFOLD GENESISEach moment, it receives only scattered readings from 3% of the surrounding volume.

3IT IMAGINES THE RESTFrom those readings, it maintains a full 32³ internal model and adapts when motion changes.

LEARNING LOOP: ON How much can it see?

1THE REAL BODYThe answer, shown to you for comparison. Genesis never receives this complete view.

1 → 802 cells
Green = sensing tissue · amber = stable core

2WHAT GENESIS SEESOnly these scattered sensor readings enter the model. Everything else is hidden.

31,785 voxels hidden
983 readings supplied

3WHAT GENESIS PREDICTSIts complete internal picture—including the parts it was never shown.

32,768 voxels predicted
Blue = the watcher's internal model

bright dots = occupied readings dim dots = empty readings Drag any view to rotate all three together.

match to real body—

sensor readings—

motion skills saved—

body movement—

What is happening now?

Press START. Genesis gets one complete starting frame; after that, 97% of the volume goes dark.

Technical readout appears here.

same 72-frame test · learning loop on81.3% final match

same test · learning loop off55.3% final match

Why this matters: without correction from sparse observations, the internal model drifts. The learning loop keeps it attached to reality.

▸ technical scope and provenance

Target provenance loads from the GPU node.

// symbolic operator synthesis — teach a grid rule, compose moves, prove transfer

COMPOSE: ON SCRAMBLE: OFF

examples supplied

challenge · committed seed —

demo mismatch · fraction of example cells still wrong

—

// continuous Cl(3,3) operator-forming — teach a transform from points, transplant to unseen points

2D field — examples (○→●) · queries (＋ predicted vs ✕ truth)

operator blades

held-out generalization

— R²

transplant (new cloud, same O) —

// composition lab — form A · form B · gp(O_B,O_A) · depth dial · fair nulls

ROTATE-ONLY: OFF

depth k = 1

composed field at depth k — ＋ predicted vs ✕ truth

depth curve · R² of O_BAᵏ

composition R² (k=1)

—

depth k R² —

blades · O_BA = gp(O_B, O_A)

fair nulls — must fail

// autonomous active inference — dual: structure vs scramble, no answer key either way

prediction error over ticks — green=structure · red=scramble

generalization R²

structure

—

scramble

—

// teach A, then B — flat bank forgets · bifurcation bank retains both

regime A · bank specialist

regime B · bank specialist

retention after teaching B