ExoKey

The bone. 410 members → 1632 curved sub-beams · 7.54 g beam / 12.7 g hollow tubes (0.8 mm wall) · every surface ≥ 1.5 mm · no spikes, no loose ends, one piece · 481 µm at the wells against a 500 µm gate · 1021 pillars, 0 props. Grown, not re-pruned — the render takes the topology the objective grew bottom-up; a top-down re-prune instead dead-ends in a heavy uniform 34 g membrane — a local optimum it cannot escape (§8.15k).

The sensor. Each well is a contactless-Hall flexure — a magnet on a printed TPU dome over a 3-axis Hall. The read-out is modelled (a keypress swings ~430 LSB, 200× the noise); the mount that holds the magnet + Hall is built entry-first — the finger's slide-in route is a hard constraint, so nothing crosses the path the finger comes in by — and to the measured pad, so the fingertip actually seats in the cup (it used to float above it). Four long fingers share one cluster; whole part 40.2 g, one watertight solid.

The strap. The band is the convex hull of (skin ∪ device), so it rides over the gauntlet it holds down; a watch-lug captures a pin in shear at each anchor foot; one adjustable strap (21 mm range) fits the 5th–95th percentile population.

The impact. A 50 N knock, not the keypress, is what sizes the structure — it breaks a deflection-only bone (348 MPa vs 70 MPa yield), and it wants a broad, load-sharing skeleton, not the sparse one thickened. Grown with the knock in the loop, co-sized for the gate and the stress, then form-found so its load paths straighten: 23.2 g, 39% lighter than bolting it on (37.7 g, §8.15k).

• entry.html — the finger-entry route, all five digits (§8.15l). For each finger the channel it slides in by — the swept fingertip — passing clear of the mount and the struts. The cup is built to the measured pad, so the finger both gets in and seats (a render caught it floating ~7 mm above its cup; now it rests on the floor, 0.4–0.8 mm).
• harness.html — the minimal-copper harness bus (§8.15l qqq-2). The sensors are I²C, so the wires need not run point-to-point: VDD/GND are shared and SDA/SCL are a bus. The harness is the exact minimum Steiner tree over the struts (Dreyfus–Wagner) — 275 mm against 490 mm of five independent runs, −44% copper. Toggle the old routing in the legend to see the sharing. (Not quite TSP — the branching tree beats the daisy-chain.)
• onstrap.html — the gauntlet on the OUTSIDE of the strap (§8.15j, the current decision). Hand → soft strap (the only thing that touches the hand) → gauntlet. The strap cushions, tethers, and spreads; no hard bit meets the hand, and the gate still holds (481 µm).
• impact.html — the impact-aware gauntlet (§8.15k). Grown WITH a 50 N knock in the load set and co-sized for the gate AND the knock: a broad, redundant skeleton (747 struts) that shares the blow, then form-found (relax_nodes) so the load paths straighten — 39% lighter than thickening the sparse keypress bone, and no longer jaggy. Each strut at the radius the gate+impact sizing chose; bright = the knock's load path.
• strap.html — the strap, over the gauntlet. The band routed as the convex hull of (skin ∪ device) so it bulges over the structure it pulls down; the gauntlet as solid tubes, and the anchor feet the watch-lugs sit on.
• typing.html — watch it type. The hand posed into every wired press: the character, the direction, the muscle effort, and whether the digit can actually balance the key. Every pose derived from the pad Jacobian, not drawn. The gauntlet is the real one, at the radius the physics chose for each member, with the wall translucent so you can see the marrow cavity — and the sacrificial supports on a toggle.

Artefacts, newest first: bone.npz (the answer — friendly, curved, hollow), gauntlet.stl (the printable solid; slice it with 2 perimeters and 0% infill — that is what makes the members hollow, and the printer was going to do it anyway), friendly.npz (the ergonomic floor, solid rods), smooth.npz (the spline study), printable.npz (the nozzle floor — the thin one nobody could bear to wear), wire.npz (the welded-wire build), sized.npz, final.npz, pareto.pkl.

Superseded renders were deleted (all regenerable from scripts/). Invalidated results were archived, not deleted — see out_archive/.