The K-lens · real CPython, computed live in the page

The K-lens, running.

Not precomputed numbers replayed — the tiny transformer's forward pass, the tangent J·δ, and the rotation across the jump are all computed in-browser by real CPython on the actual embedded weights. Change direction or bracket width; it recomputes for real. LIVE

⚠ fetches the Pyodide runtime (~6 MB WASM) from cdn.jsdelivr.net on load. NOT offline. Needs network once.

booting CPython…

direction δ:

K bracket gap (from A and B): 0.01

what the lens sees, reading down the stack — the words it leans toward, the shape of its distribution, and the dead space where nothing's decided (auto-plays layer by layer):

your input:

layer: L0

what the K-lens actually sees — the two tangent vectors J·δ (near A, near B) in the plane they span, and the arc between them:

the raw 16 components of each tangent (this is the vector the lens holds):

—

K-lens rotation (live)

—

full A→B tangent turn

—

compute time (ms)

—

runtime

what runs in-page: hand-rolled numpy forward pass of the exact from-scratch transformer (weights embedded), tangent by central finite-difference JVP, rotation = angle between the two bracket tangents. verified self-consistent to 0.01° across four decades of FD-eps.
honest gap: the in-page numpy uses FD tangents (torch used exact autograd) — within ~1° of the torch reference; the method is identical, the discrepancy is FD-vs-autograd, disclosed.
Möbius: chosen geometry for "a frame that rotates as it rides the path" — not a topology claim. the twist you see = the rotation the Python computed.