Order 0: pre-registration¶
The point of Order 0 is a public comparison of measured boards with simulations that were published before the boards were made. This page says what will be frozen, how the freeze is made verifiable, and how the data will be judged. It is a draft for the owner’s review: the criteria below become binding when the pre-registration release is published.
What is frozen before upload¶
The release (a GitHub release of this directory, its tag order0-prereg-<upload>) holds, for
each upload:
The boards: the KiCad projects and the sha256 of the Gerber zip that is uploaded to OSH Park (from the fab bundle’s manifest), the stick catalogue, the generator commit. The sha256 is taken from the bundle built from the committed board files: regenerating them with another Python or numpy can move a coordinate by 0.1 µm, so the release names the interpreter (
bazel run //yapnr/rf/coupons:cli, the pinnedrequirements.lock). A board whose optimizer window is still empty failsyapnr fab check(FAB-PLACEHOLDER) and cannot be built.Predicted Touchstone files for every stick, at nominal FR408HR and at nominal EM528:
coupons (lines, width set, masked line, ring, stub, switch-term coupon, C-pads): the coupon forward model, a 2D quasi-static RLGC cascade (predictions/), on a 30 MHz grid; the numbers the criteria test are pre-registered as scalars besides (
scalars.json: the ring’s and the stub’s notch frequencies to 10 kHz;lines.csv: every line family’s εeff, α and Zc on the 6 MHz measurement grid);R1, R1t: the RF optimizer’s FDTD, forward only, on the same substrates as the demos, at the solver’s second grid refinement (0.05 mm on M, 0.25 mm on W), on which every edge of their copper lies (R1’s 0.70 mm arm is 14 cells about a node there; on the 0.10 mm grid it would be drawn 0.6 or 0.8 mm). Their geometry is frozen in this release as drawn: a quarter wave from Hammerstad’s T-junction reference plane, a closed-form correction; it is not tuned after any D1 or D2 result;
the ring A11 and the stub A12 also in the optimizer’s FDTD, forward only, with R1’s settings (0.05 mm, the FR408HR and EM528 thickness-equivalent substrates): a second, independent prediction of the held-out notches next to the coupon model’s;
D1, D2: the optimizer’s validation runs on the thickness-equivalent and on the nominal substrate (and on EM528’s thickness-equivalent one), with the spec,
result.json,validation.jsonand the footprint;the thru and the reflects (A01, A06, B01, B05) are calibration standards (the thru is the identity between its reference planes, the reflect is not assumed), A15 has no RF port; the tier-1 view of A01 and A04 (the launches included) is the openEMS prediction below. Every solver prediction names its source: yapnr commit
917222a(the bundle the cloud tasks ran; theresult.jsonprovenance shows the image wheel’s version string,688fac3) and the imageghcr.io/studio-fug/yapnr@sha256:f41a4760…with its native float64 kernel.
The solver’s loss correction: the predicted |S21| of the demos and references is published twice, as the solver gives it and corrected for the conductor loss the solver does not model (rough copper, the ground plane’s loss, ENIG); the corrected one is the prediction. The correction makes the solver’s α(f) of a straight 0.40 mm (M) or 3.0 mm (W) line equal to the shipped coupon model’s on the same nominal stackup (
lines.csv; M 0.132 dB/cm at 5 GHz, with the roughness prior Rq 1.0 ± 0.5 µm). Run 0b measures the solver’s α(f) from straight lines of two lengths on the references’ grid (0.40 and 0.70 mm wide, 10 and 30 mm long); the difference Δα(f) is applied along R1’s path (the 0.70 mm arm to the junction, then half the output line), and for any geometry (D1, D2) the fraction of the incident power the solver dissipates is scaled by α_coupon/α_solver of the port line (demos). Closed forms disagree on the smooth part (the coupon model’s 2D Wheeler factor gives 0.037 dB/cm on M at 5 GHz, Hammerstad-Jensen 0.059), so the measured α(f) of the lines, not either model, settles it, and the post-fit prediction uses the fitted loss. Each substrate is corrected with its own lines (M-eq: run 0b’s; M-nom, M-eq-em528, W-eq, W-nom, W-eq-em528: the same lines run on that substrate) against the coupon model of the board it stands for (FR408HR, or EM528 for the-em528substrates); R1t uses the 3.0 mm line’s Δα along its path (region W has no 5.0 mm line run). The corrected files and the raw and corrected worst cases are in predictions/corrected/ (python -m yapnr.rf.order0 correct).The full optimizer history: every D1 and D2 attempt with its validation, failures included, and the rule that picked the shipped run.
An independent 3D prediction (openEMS 0.37, predictions/openems/): D1 and R1 with 43 µm copper on the nominal FR408HR stack at two meshes, the 0.40 mm line, and the sticks A01 and A04 with the Cinch launches as the tier-1 calibration sees them, with the comparison against the optimizer’s FDTD stated as found. This was originally run against part 2’s D1 export (−17.4/−18.2 dB |S11| at 0.05/0.025 mm, a miss, against yapnr.rf’s −20.9 dB on M-eq); that export was superseded 2026-10-05 by
d1-starre-optimized under PR #53’s fixed corner-gap DRC (label7e070ca8, predictions/D1-d1c), and openEMS was re-run against the new export (0.05 mm mesh only; the 0.025 mm convergence pair was not repeated, for budget). Like for like, against yapnr.rf on the same nominal substrate (M-nom): the two agree on |S21| within 0.14 dB, but openEMS predicts a 2.9 dB miss of the −20 dB |S11| spec (−17.10 dB) while yapnr.rf puts D1 at the spec line (−19.70 / −20.02 / −20.05 dB coarse / fine / finer; −20.77 dB on its M-eq design substrate), a 2.6-3.0 dB disagreement, essentially the same as on the old export. A third solver, Palace (independent FEM), has not been run for D1 or R1 in either round: an open item, not silently dropped. Both solvers’ predictions are registered as they stand, misses included.The criteria (below) and the analysis order.
D1 and D2 carry the first 8 hex digits of the sha256 of their result.json in silkscreen, which
ties the physical part to its prediction (D1: 7e070ca8, superseding part 2’s ad20e643 after
the corner-gap DRC fix; D2: 03b7d938).
Proof that the predictions came first¶
A commit date proves nothing (authors set it). Before upload:
the release is created, and the sha256 of its bundle is stamped with OpenTimestamps (free); the
.otsproof is added here;optionally, the release gets a Zenodo DOI.
Acceptance criteria (draft)¶
The table is the Order 0 design’s (§7), with its review’s replacement of “agrees with prediction” (the −15 dB band edges it first named lie above the LibreVNA’s 6 GHz: R1’s −15 dB band is 3.3-6.7 GHz, 68 %). D3 and D4 (the antenna and the diplexer) are stretch items that do not ship on Order 0; their rows are kept as written.
Item |
Criteria |
|---|---|
Coupons, held out |
Ring f_n and stub notch within 0.5 % of the post-fit model. Verification line consistent with the mTRL uncertainty. |
D1, D2: “meets spec” |
Measured (mTRL-corrected, renormalized to 50 Ω) |S11| ≤ −20 dB and |S21|, |S31| ≥ −3.4 dB over 4.25-5.75 GHz |
D1, D2: “meets validation criteria” |
|S11| ≤ −15 dB and ≥ −3.6 dB |
D1, D2: “agrees with prediction” |
Max over 4.25-5.75 GHz of the vector difference |S11,meas − S11,pred| ≤ 0.05 (a −20 dB S11 is 0.10), in the mTRL native reference (line Z0, the same as the solver’s modal port); frequencies of in-band |S11| minima within ±1.5 %; |S21| within ±0.25 dB; |S21| − |S31| within 0.3 dB. “Agrees” also counts a difference that lies inside the k = 2 measurement uncertainty, which is reported next to each number. |
D1, D2, R1, R1t: reference impedance |
“Meets spec” is evaluated after renormalizing to 50 Ω with the fitted line Z0, and that Z0 uncertainty is propagated. ±1.3 Ω without A16 moves a −20 dB |S11| by about ±1 dB; ±0.5 Ω with A16 moves it by ±0.4 dB. |
D1 against R1 |
The claim tested is “the optimizer automatically matches a textbook-class design”, not that it beats one. Report both against the same criteria, plus footprint area, worst in-band |S11| and |S21|, imbalance, compute hours and the number of optimizer attempts. |
D3 (stretch, not on Order 0) |
Frequency of the |S11| minimum within ±1 %, and the −10 dB bandwidth within ±20 % (relative) |
D4 (stretch, not on Order 0) |
Channel centres within ±1 %, insertion loss within 0.3 dB, rejection within 3 dB of prediction while at least 10 dB above the floor |
Each result is reported three ways: against the pre-registered nominal prediction, against the post-fit prediction (stackup fitted from the coupons), and against the spec. The instrument floor is drawn on every transmission plot.
How the rows are applied (decided before any measurement; none of these changes a limit):
The prediction compared is the loss-corrected one (predictions/corrected/): the solver’s |S21|, |S31| with the conductor loss it does not model added from the coupon model, per substrate. “Meets spec” uses −3.4 dB (run 0b, D-O0-11: the corrected R1 reaches −3.285 dB, demos); it applies to R1 and R1t as well, which are judged like D1 and D2.
The reference of “agrees”: the solver’s modal port is calibrated on the open-L1 line of the run (
yapnr.rf.ports.calibrate_line, recorded with each result); the multiline TRL gives the measured data in the fitted Zc of the board’s line, which on W has its L1 pour and the inner-layer ring (about 48.6 Ω against the solver’s 50.0 Ω, |Γ| 0.014 per port). The measured data are therefore renormalized from the fitted TRL-line Zc to the run’s port impedance before the comparison; on M the two agree within the fit’s uncertainty.Measured notch frequencies (ring, stub) come from fitting the model’s notch shape to the data around the notch, not from the lowest sample (the n = 3 notch lies 10-20 dB above the instrument floor).
D2 ships only as the owner decides (no D2 formulation meets the −20 dB |S11| spec; the nearest,
d2-star-sched, reaches −19.3 to −19.4 dB and passes its |S21|, |S31| validation limits, D2). If it ships, it is judged by the same rows, labelled a design that missed its spec (see below: |S11|, and loss-corrected |S21| on FR408HR).Owner default (recorded by the main loop, 2026-10-04):
d2-star-schedships on O0-W as this pre-registered near-miss unless the owner says otherwise. What its predictions say (yapnr.rf only; no independent solver has run on D2) [S]:|S11|: −19.3 to −19.4 dB on W-eq and W-nom (coarse/fine/finer), a 0.6-0.7 dB miss of the −20 dB spec; −19.9 to −20.1 dB on W-eq-em528.
|S21| = |S31|, raw: inside the design’s validation limits (−3.45 dB coarse, −3.6 dB fine/finer) everywhere, but the committed in-job criteria (inputs/criteria/d2.json: −3.40 dB coarse) fail on the coarse grid: −3.4117 dB (W-eq), −3.4205 dB (W-eq, wide pulse), −3.4334 dB (W-nom). Both versions are reported.
|S21| = |S31|, loss-corrected (the prediction “meets spec” uses): −3.43 to −3.45 dB on W-eq and W-nom, a 0.03-0.05 dB miss of the −3.4 dB spec; −3.35 to −3.36 dB on W-eq-em528 (predictions/corrected/).
So on FR408HR D2 is predicted to miss the spec on |S11| and, by a smaller margin, on loss-corrected |S21|/|S31|; it is judged by the same rows and labelled a design that missed its spec, not only its |S11| spec. The board carries the first 8 hex of its
result.jsonsha256 (03b7d938, back silkscreen “O0-W D2 03b7d938”) and a QR label stick (B15) to its predictions page; its run id and fab-DRC status are in O0-W’scatalog.json(catalog.o_optimized), not in silkscreen. This default is a placeholder for the owner’s own sign-off, not a substitute for it: the row above (“ships only as the owner decides”) is the binding text until the owner confirms or overrides it.
Criteria are never changed after data arrive; if one must be, results are reported against both versions.
The k = 2 uncertainty model¶
Every comparison above carries a per-frequency k = 2 band (see Uncertainty to publish): the linear propagation of the multiline TRL’s inputs (Hatab’s method and code, uncertainty-multiline-trl-calibration, BSD-3: standard definitions, line lengths, noise), the measured connection repeatability (three connections), the copy-to-copy spread (two copies, one lot), the fitted line Z0’s uncertainty and the multiline-TRL against IEEE 370 2x-thru difference. Before the boards exist, the expected size is the Monte Carlo of the coupon set with the LibreVNA’s specified noise and a prototype connection-repeatability model (solder variation σ_L 0.02 nH, σ_C 0.01 pF per launch): εeff ±0.37 % per point (±11 MHz at 5.8 GHz), α ±0.0044 dB/cm, Z0 ±1-1.5 Ω model-assisted (±0.5 Ω with A16); twice the solder variation gives εeff ±0.65 %. The measured repeatability replaces the prototype numbers.
The C-pads (A16): how h is read¶
A16’s two pads (6 and 12 mm square) give the prepreg height h, and with it the absolute Z0. The launch and the 10 mm line before each pad are not negligible at low frequency (about 10° at 300 MHz, against 0.4° for 1 % of the 12 mm pad’s capacitance), so the pad-pair difference does not cancel them. The procedure:
De-embed each pad’s 1-port with the launch half of the thru A01 (IEEE 370 NZC, the same tier-1 data); the mTRL error box is the check above 0.5 GHz.
Fit 0.1-1 GHz with the distributed pad model (each pad an open-ended wide line with its open-end extension,
predictions/.../A16-*.s1p), h and the prepreg’s εr free, the line-to-pad step included as its closed-form shunt capacitance.Uncertainty budget: besides the calibration, the dielectric’s dispersion (Djordjevic-Sarkar puts εr at 100 MHz about 2 % above 5 GHz; with tan δ uncertain by 0.003 that is about ±0.7 % on h) and the step model (about 0.5 % of C12 − C6), next to the ±1.6 % the design expects.
Analysis order (blind)¶
At capture, each demo’s raw files are hashed and committed, not de-embedded or plotted.
The coupon fit (
yapnr-stackup-fit/1) is committed first.Only then are D1, R1, D2 and R1t de-embedded and compared. The post-fit prediction never sees demo data.
Uncertainty to publish¶
Per-frequency k = 2 bands combining: the linear propagation of the calibration (standard definitions, line lengths, noise), the measured connection repeatability (three connections of A01, A03, A04 and each demo), the copy-to-copy spread (two copies, one lot), the line Z0 uncertainty, and the difference between multiline TRL and IEEE 370 2x-thru de-embedding from A01. Not covered, and said so: one lot, two assembled copies, an instrument with preliminary specifications.