Spacetime sheet fused with a memristor · the excitron latches

The memristor basin.

The sheet on top is the flat field — the tangent. The well below is the substrate's measured curvature: a real, anisotropic second-order landscape (not an isotropic bump). The excitron is a test charge that falls into it and latches — and it carries a real memristor state w that freezes a weight. A small kick is restored; a reverse write-pulse re-writes it. The Jacobi lens reads that frozen echo, not the live fall.

LIT — the well is a MEASURED 2nd-order curvature (Hessian xᵀHx), anisotropic ≈6:1 — not isotropic κr² LIT-MATH — the memristor state runs a real Strukov ODE; its pinched-hysteresis invariants are verified live (break one → INVALID) SPEC — that a transformer literally freezes a weight this way is a proposed reading, not a measured property of the model
device stack: [ anode — the flat field/sheet │ excitron — the test charge + its memristor state w │ substrate — the measured basin (single well) │ cathode — the floor where w freezes ]
drag = orbit · scroll = zoom
the memristor, running · i–v pinched loop · invariants: VALID
drive ω
the well: the surface is the substrate's measured second-order curvature — a real, anisotropic (≈6:1) landscape, a single basin. It is built once from measured data and is not deformed by the excitron: the excitron is a probe falling in, not a mass bending the sheet. The bottom of the well is where the state freezes.
the state (real): each excitron carries a memristor state w∈[0,1] stepped by a Strukov/HP ODE (dw/dt = η·μ·Rₒₙ/D² · i, clamped). A SET pulse drives w→1, RESET drives w→0, a small read-kick perturbs position but w is retained — read-after-write shows RESTORED. Since the terrain is one measured well, the memory lives in w, not in a second basin to escape to.
the proof (LIT-math): the panel above runs that ODE live and draws its i–v curve — a real pinched hysteresis loop. Two invariants are checked each frame: (a) it passes through the origin (at v=0, |i|<ε for every amplitude — Chua's defining test), (b) the loop area shrinks as ω rises. The badge is green only while both hold; press break the pinch to inject an offset and watch it flip INVALID — the control arm.
honest: the well is real measured curvature; the memristor math is real and verified here. What stays SPEC is only the reading that a transformer literally freezes a weight this way — a proposed answer to "what's stored in the gap," not a measured property of the model. real WebGL, offline, fails loud (including lost-context).
provenance: well: measured Hessian field · loading…