Tutorial › reference

• Nothing is authored
• The model, and what it costs
• The technology file
• The collaborator's technology rules
• What a device becomes
• What it found
• The fab view
• Counted versus declared is a disclosure, not a verdict
• Integers, and one implementation
• Out of scope, deliberately

The electrodes a device implies#

The layer that turns a trap graph into metal with coordinates, and metal into an ion height. Before it, the only "electrodes" anywhere in this project were the authored integer control.wiring.electrodes_per_trap = 24 and a decorative pill tiling in the browser whose count came from drawn pixel length. Neither was falsifiable. Now rectangles exist, they have coordinates in nanometres, a fab tool opens them, and a field solver reads them.

The browser page can carry the metal as a backdrop -- qccd phys <device> --html -- but it is a view, not an editor: see the fab view for what it is deliberately not.

python -m qccd phys ring144_24v                    # the metal, and its DRC
python -m qccd phys chain --svg chain.svg
python -m qccd gds ring144_24v -o ring.gds         # 1372 polygons, database unit 1e-9 m
python -m qccd sweep field chain rf.w_rf_nm 60000,99500,140000 --set mass_u=137.905247

Nothing is authored#

build_layout(arch, tech) is a pure function. No field was added to Node or Segment, no section to .arch.json, no SCHEMA_VERSION bump, no architecture file changed, and the browser is untouched. The metal carries no information the pair does not already determine, so storing it would only create something to lose — Knowledge: d_technology_is_a_sidecar.

The model, and what it costs#

Planar electrodes in the z = 0 plane, ion above. Two assumptions, taken verbatim from the paper this layer is checked against:

1. that there is no gap between electrodes;
2. that the entire plane is covered by conducting electrodes.

— arXiv:2201.12579 ms.tex:200-208

Each electrode then contributes the solid angle it subtends, over 2π (Wesenberg, arXiv:0808.1623, eq. biotsavartpot; the companion analytic treatment is House, PRA 78 033402, which has no arXiv preprint and is recorded in a ledger of non-arXiv sources). No mesh, no BEM, no external solver, no new dependency.

Every tolerance in this layer is quoted from one table, the same paper's measurement of its own model against FEM of the fabricated trap:

Wesenberg also states when the model may be used, and one condition drives a knob:

"necessary conditions … include that gaps between electrodes are much smaller than d, that the extent of the trap is much larger than d, and that the distance from ion to other conducting surfaces is much larger than d"

A finite rail violates the second in proportion to (h/L)². That is measured, not assumed: the truncation is monotone and second order, and rail_length_over_h exists because of it.

The technology file#

A sidecar, never part of the document. Dim has two fields and both are required — there is no Dim(41500). A source beginning declared: marks a number this project chose rather than read, and Technology.declared() lists them.

The one shipped preset has none. Every dimension of eth_junction_2201.12579.tech.json is either a page reference or a derivation checked against its inputs:

Two lattice scales, and why one is not enough#

nm_per_unit_x = 225000 nm: one axial trap pitch, which is three broadcast control electrodes. arXiv:2305.03828 states the conveyor belt supports "20 wells on each side (one for every three electrodes)"; 2201.12579 gives 75 µm segments. 3 × 75 = 225.

nm_per_unit_y = 355000 nm: one min_axis_pitch, the closest two parallel trap axes can be drawn, because each needs 175 µm of electrode stack on the facing side plus a gap.

These measure different things and they disagree. An isotropic 225 µm was tried first and shorts the north control column of every ring device against the opposite rail — the two rows physically do not fit.

The collaborator's technology rules#

A reviewer of the hardware model asked for six named variables, each with a minimum and a default. They are first-class fields of the technology, checked when a file loads, and a breach is a TechnologyError naming the field, the value and the minimum. Three of them are dimensions this layer already had under other names, so the rule reads the existing dimension instead of storing a second copy that could drift.

Two rules are about the drawn metal rather than about a number in the file, so they are DRC rules and nothing else:

• overlap — any two electrodes, on any layers, whose shapes share positive area. One finding per pair of nets, with both net ids in the message. Same-net shapes are skipped, because two perpendicular RF rails necessarily overlap at every degree-4 node and on one net that is a merge; touching at 0 nm is a clearance finding, not an overlap.
• dc_pairs_per_site — for every trapping site, the number of pairs (a north and a south pad at one axial position) centred within the site's span, against n_dc_pairs. A site's span is one trap pitch of rail positioned where the site's rail is: half a segment either side for an interior site, and for a dead end the window is shifted onto the side that exists rather than shrunk. Only pads owned by the segments incident to the site count, so a perpendicular rail's electrodes are never somebody else's.

w_rf, w_dc and l_dc reach the drawn metal through min_width, which measures against the process rule of the layer, not against the technology's design value: a technology that declares 50 µm electrodes and draws 5 must fail somewhere, and it fails there.

Electrodes may be shapes other than rectangles — and are not yet. The collaborator asked for triangular or wedge corner electrodes around a junction so that nothing has to overlap. The union by net is exact for rectangles only and refuses any other polygon rather than growing it to its bounding box, and every gap and overlap computation in the design-rule check goes through that union. A wedge would therefore need a new gap computation before it could be drawn, which is more than this change. Junction pads are omitted where they would overlap instead (keepout_half_width), the omission is counted in the layout notes, and polygon corner electrodes are recorded as the intended follow-up.

A minimum can be waived, by name, in writing#

eth_junction_2201.12579 reproduces a trap its authors fabricated at a single 5 µm gap (ms.tex:368) — between control electrodes and between an RF rail and the control metal beside it alike. The project minimum for g_rf is 8 µm. Applying it here would move dc_setback from 125.25 µm to 128.25 µm and redraw every rail-to-column clearance of a published geometry, reporting a trap that paper never had.

So the preset carries a waivers block naming g_rf and saying why, and that is the only way a number under a minimum loads at all. The rule is still checked — it was waived, not switched off — the waiver is listed by Technology.waived(), printed in the layout notes, and printed by the DRC as a waived technology minimums disclosure. A waiver for a rule the technology already meets is refused too: a stale waiver is a claim nobody checked. The published preset's geometry is therefore byte-identical to what it was, which is the property that matters for a reproduction.

The second preset: surface_default#

python -m qccd phys ring144_24v --tech surface_default
python -m qccd gds chain --tech surface_default -o chain.gds

The rules at their defaults rather than their minimums, with every other dimension derived from them and the arithmetic in each source string. Every number is declared: — it is this project's own reference process, not a reproduction, and declared() is the query that tells the two presets apart.

The two scales are equal, which the published preset cannot manage. The axial unit is raised to the transverse floor instead of the two disagreeing, so one lattice unit means one length in both directions. Two floors and one quantum fix it: parallel axes need min_axis_pitch = 318 µm; a dock spur's rail overrun needs 2(dc_setback + w_dc + rail_end_extension + g_rf) = 446 µm; and the scale is a whole number of pitches so the tiling is uniform across every site. 8 × 58 = 464 µm is the first multiple above both. The cost is eight control electrodes per trap pitch where three would do — and the benefit is that ring144_24v fits: 44 rf_dc_clearance failures and 44 overlap failures become 0.

Every site's span carries n_dc_pairs pads#

_tiled_starts lays the n_dc_pairs pads nearest a site from that site's end of the span and leaves any slack in the middle of the rail, so the count under a trap is the technology's number and not a remainder. Below 2·n_dc_pairs pads there is no middle to put slack in and the run is simply centred — which is what every segment of the published preset does, three pads in a span that holds exactly three, so its metal did not move.

A pad is beside a site, not centred on it: a pad centred on a node would straddle two segment cells, and flatten would qualify the two halves with different owners — two electrodes abutting at 0 nm, which the DRC would then report as a short. Moving a trapping position onto a pad centre needs the tiling to be owned by the nodes rather than by the segments, and that is a larger change than this one.

What the rule found#

Under the published preset a lattice unit is three control electrodes long, so a single pad dropped at a corner already starves the site beside it — five devices that were "clean" are not, and what changed is the question being asked, not the metal. Under the defaults most of that goes away and what is left is architectural: ring144_24v's 24 docks hang half a lattice unit off the rail, inside the keep-out where a crossing rail's own control column sits, so they get one pair and no technology that keeps pads off that column can give them three. grid9x9 and deck_unit_cell keep 144 each because their sites are one lattice unit from a junction in every direction.

What a device becomes#

Each axis-aligned segment gets two RF rails flanking its axis at w_g/2, one segmented centre control electrode between them, and two segmented control columns outboard at dc_setback. A non-axis-aligned segment is refused by id, never approximated. All 1,392 segments of all nine shipped devices are axis-aligned, and no lattice coordinate needs rounding — measured, not assumed.

Two things are interrupted, for the same reason:

• a control pad is dropped where a perpendicular rail crosses its column;
• a rail stops at a perpendicular trap's gap.

The second is not obvious and is not what the build plan said. Running both rail pairs straight through each other lays RF metal across both trap axes, and solving that finds no confined ion position anywhere near the junction — a 175 µm dead zone, because the ion would be directly over driven metal. Interrupting the rails leaves the four L-shaped quadrant electrodes of 2201.12579 fig. 7(a), which is the geometry that reproduces its numbers. Knowledge: fd_the_naive_crossing_must_keep_both_gaps_clear.

The junction is the paper's counterexample, on purpose#

A degree ≥ 3 node is drawn as the naive crossing and tagged role='naive_crossing', with the citation travelling in the layout notes. The optimized alternative in that paper is a cubic B-spline boundary fitted by Nelder–Mead; shipping it means shipping an optimizer. Drawing the counterexample costs nothing and yields a checkable number.

What it found#

Numbers this platform could not previously state. Each is pinned by a test.

The naive junction is as bad as the paper says. At 2201.12579's own conditions (⁴⁰Ca⁺, 40 V peak, 2π × 40 MHz), fed only the two published RF widths:

Nothing is fitted, and the assertions are written to the paper's bands and no tighter: the 84 µm is read off a figure. Far out along an arm the same geometry returns the linear-section closed form to 3.6e-4, approaching it monotonically, so agreement at the centre is not a coincidence of arm length. PLAN §0.5's prose about RF barriers is now two checked numbers.

No shipped device sits at its design height. A rail sized for 49.95 µm delivers it only in isolation.

Four of the nine are outside the model's own 5% band, so the shift is geometry rather than numerics. The null also moves off the trap axis, which is why height is solved in the transverse plane (transverse_null) and not in z alone — a search constrained to the axis reports no null at all for eight of the nine.

ring144_24v is not fabricable in this technology, and the threshold is exact:

along the rail    (w_g/2 + w_rf) + (dc_setback + dc_width) + gap = 300250 nm
across the rails  2 (dc_setback + dc_width + rail_end_extension + gap) = 460000 nm

The shipped scales are 225000 and 355000. One nanometre under either threshold and the violations return. The axial trap pitch the control electrodes imply is 225 µm and the axial requirement is 300 µm; they are incompatible. That is an independent geometric argument for PLAN §0.5's conclusion about the 24 verticals: they do not merely cost quanta, at these dimensions they do not fit.

The naive crossing also deletes control electrodes — 456 on ring144_24v, and all of them on grid9x9 and deck_unit_cell, which as drawn have nothing left to control them with. Unlike the confinement this is a count, not a simulation.

Four segments in cyclone_dual_loop run through nodes they are not incident to. EA35 goes from (35,0) to (35,3) straight through DT35 and DB35, both degree-2 nodes on the data loop. A planar trap has no overpass, so the graph and the plane disagree: if the crossing is real those nodes are degree 4, they are junctions, they are not being charged as junctions, and the router may send two ions through the same metal. ring144_24v has two more at degree 1 — its end-cap docks, which the architecture reference already records. Reported, never repaired: the architecture is not this layer's to edit, and the two possible repairs say different things about the machine.

The fab view#

qccd phys <device> --html out/x.html renders the studio page with the derived electrodes underneath the schematic, in a gMetal group that sits first in the SVG and therefore below everything else.

It carries its own transform, and the page does no arithmetic to get it. The scale, the fit and even the scale bar's rectangle are computed in Python by qccd.phys.svg.metal_view_model and read verbatim. They have to be: the page's own px()/py() is anisotropic by up to K_ANISO = 12, and pushing a 99.5 µm by 16 mm rail through it would draw a shape no fab could make. A test strips the block's comments and strings and asserts it contains no *, /, Math., px( or py(.

The underlay does not register with the schematic, and the page says so. Registration would need the page's sx/sy to equal the technology's nm_per_unit_x / nm_per_unit_y; on chain72 those are 1.0 and 0.634, and no shipped device matches. One of the two views has to misstate a proportion and it is not the one with nanometres in it, so the metal is drawn true to scale with a scale bar — the one length on the page that means a physical distance — and a caption saying which view is which.

A page that did not ask for metal is unchanged. The payload is absent from the view model unless a technology was named, so every page emitted before this existed is byte-identical to the one emitted now. The renderer never imports this layer: the metal arrives as a dict, because a page is a property of a run and the metal is a property of (device, technology), and wiring one to the other in the renderer would make every page pay for a field solve it did not ask for.

What it is deliberately not: no browser edit, no parity bucket, no editing verb, no hit-testing, no palette tile. The mirrored half of this project is diffed at tolerance zero, and a fab view that computed anything would have to be diffed too.

Counted versus declared is a disclosure, not a verdict#

The hardware price list carries electrodes_per_trap = 24; this layer draws 5.89 per site on ring144_24v. Neither number is measured — the 24 is an authored integer that may count shim and compensation electrodes this layer does not draw, and the 5.89 follows from a dc_pitch and a lattice scale that are the technology's claims. So the report prints both, prints the axial trap pitch each implies (600 µm against 225 µm), and stops. Turning that into a pass/fail would be inventing a fact out of two conventions.

Nothing here is a verifier rule. RULE_STATEMENTS is still 23, BROWSER_SET still 17. Those are mirrored in the browser engine and diffed at tolerance zero over every architecture file, so a Python-only rule firing there is an automatic red harness — and a design rule is a different kind of claim anyway, about a technology's fabrication limits, which the document does not declare and the browser cannot know.

Integers, and one implementation#

The shape table and the technology contain no float literal — checked by tokenising, so a comment saying "0.83h" is fine and 1e3 cannot sneak through. Coordinates are integer nanometres from the lattice to the GDSII file, whose database unit is 1e-9 m, so the round trip is == on lists of ints rather than a tolerance. Placement is quarter-turns through one point transform; four of them compose to the identity by equality, not by approx.

The field kernel is Python only, forever. A transcendental solver cannot be differentialled at tolerance zero — atan2 is not required to agree to the ulp between CPython and V8 — so it is not written twice, and a test asserts that no symbol of it reaches the browser.

Out of scope, deliberately#

No DC voltage solution, no waveform synthesis, no axial confinement, no gap modelling, no multi-layer solve (a non-zero z_offset refuses rather than approximating), no heating rate from geometry, no optimized junction geometry, and no trap depth — a 1-D bisection above the null returns a local maximum on one line, not the escape saddle, and depth is the quantity most sensitive to every declared hole. hardware_report.area_mm2 still reports 0.0; connecting it is a separate decision.