D2: the optimizer’s thick divider¶
Compute stage, 2026-10-04 (corrected on review). No formulation meets the −20 dB |S11|
production spec (8 formulations in two waves, runs/, selection.json;
wave 2: inputs/wave2/); the nearest, d2-star-sched, reaches worst
|S11| ≈ −19.3 to −19.4 dB (coarse/fine/finer), a 0.6–0.7 dB miss.
|S21|/|S31| validation, corrected: an earlier draft of this page checked d2-star-sched
against −3.40 dB (coarse) / −3.55 dB (fine, finer), figures with no recorded decision behind
them, and reported a 0.012 dB miss at the coarse grid. The design’s §4.4 fixed validation limits
are actually −3.45 dB (coarse) / −3.6 dB (fine, finer), unchanged by D-O0-11 (which sets only
the −3.4/−3.5 dB production spec, not this separate validation margin). Against the plan’s own
limits, d2-star-sched passes |S21|/|S31| at every grid (coarse −3.412, fine −3.397,
finer −3.394 dB, all above −3.45/−3.6 respectively). So D2’s only real miss is |S11|, not
|S21|; R1t reaches −3.549 dB on the same substrate. Whether O0-W carries d2-star-sched
(re-validated on W-eq, W-nom and W-eq-em528 in variants/) as a pre-registered
|S11|-missing design is the owner’s decision; recorded here as compute.md’s recommendation
absent other owner direction.
Region W (L1 over B.Cu, inner layers removed), window 20 × 24 mm, ports W / N / S, 4.25-5.75 GHz
(Order 0 design §4.4). Its specs (one per formulation), criteria and substrate are in
inputs/ and explained in demos. If D2 ships, this directory
receives its spec, result.json, validation.json, footprint and
predicted Touchstone files (FR408HR and EM528), and its copper goes into the D2 window of upload
O0-W.
What the board around the window is, for the spec (divider-osh-w), which the solver’s open L1
over an infinite ground does not see:
the window’s west edge is P1’s port plane; P2 and P3 sit on its north and south edges 10.0 mm from the west edge (the spec’s ports); each port is a 3.0 mm feed (6 cells of 0.50 mm) running 4.0 mm straight to its launch’s reference plane;
the L1 ground pour stays 4.2 mm (3 h) from the window and the feeds; In1.Cu and In2.Cu are removed under the window, its keep-away and the feeds, and remain as a stitched ring beyond;
the reference plane (B.Cu, L4) ends at the stick’s east edge, 6.0 mm (4.3 h) beyond the window; to the west, north and south it extends 14 mm (the feeds and launches);
the TRL line of region W (B01-B05) is 48.6 Ω with that pour and ring (2D), about 1.4 Ω under the solver’s open-L1 port: the comparison renormalizes the measurement to the solver’s port impedance (pre-registration). The same holds for R1t;
the window carries
d2-star-sched’s exported copper (catalog.o_optimized; the board’s copper equalsruns/d2-star-sched/footprint.kicad_mod, 122.5 mm², polygon symmetric difference 0), with03b7d938on the back silkscreen.
Against the spec, not only the validation limits: the loss-corrected |S21| = |S31| that “meets spec” uses is −3.43 to −3.45 dB on W-eq and W-nom (a 0.03-0.05 dB miss of −3.4 dB) and −3.35 to −3.36 dB on W-eq-em528 (corrected/); the committed in-job criteria (d2.json, −3.40 dB coarse) also fail at the coarse grid. So on FR408HR the prediction misses the spec on |S11| and, narrowly, on |S21|; both versions of the |S21| limit are reported in the pre-registration.