A 42-prim Ag₂S memristive attractor-field classifier. Not Newtonian gravity — an analog attractor field: a charged probe drifts through a landscape of superposed wells and settles in the nearest one. Where it settles is the answer. A physical winner-take-all, computed by descent.
This is the device's mechanism, at reduced scale. Each Ag₂S memristive node is a well in a shared potential; an input is a charge injected at a point; the array relaxes and the probe rolls to the deepest nearby attractor — an analog nearest-attractor / winner-take-all decision made by physics, not by a clock stepping through comparisons. The real kernel places 42 such prims (the 42 nearest the origin on a hex lattice, distance-sorted) plus a 4-node CPG clock; here a handful stand in so the descent is legible.
Real: the mechanism is genuine — an analog attractor / winner-take-all computed by overdamped relaxation in a memristive potential is real in-memory computing (Ag₂S electrochemical-metallization cells are a real device family), and the canvas above computes the actual descent. There is a full C implementation, a bill of materials, and protoboard layout diagrams in the repo.
A design / disclosure, not a fabricated device: the 1 mm die, 96 µW and 100 µs are simulated targets, not measured silicon (96 µW is the board budget; the memristive array alone is ~0.4–30 µW — see the power calc in gravity.c). Retention is ~ms, not non-volatile (a 0.65 eV barrier; true non-volatility needs a >0.9 eV trap — gravity.c flags this itself). "42" is distance-sorted, not a closed hex shell. "Gravity" is a metaphor — an analog attractor field, not Newtonian gravity.
To disclose a design is real work; it is not a granted patent or a fabricated device. The interactive companion CROSSBAR · WINNER-TAKE-ALL shows the same computation as a memristor crossbar (the circuit view); this page is the attractor-field view + the real design.