RSFQ logic

sfq_predict

Margin prediction for one RSFQ cell or raw netlist: functional verdict plus bias/Ic band factors.

Fidelity

surrogate_of_josim

Validation

reference

JoSIM margin characterization under the FED drive (current-triggered feeder, reference feeder F0), which is what a real SFQ cell sees. The previous corpus used an ideal-voltage source that forces a 2*pi phase advance regardless of bias and censored the lower edge of every margin; it was regenerated on 2026-08-20, 4800 designs per cell, zero censored.

coverage

library_cells

jtl, splitter, merger, dff, and2, or2, dro, clock_splitter

fitted_model

dff, jtl, splitter

functional_and_margins

dff

functional_and_margins_spec_path_only

jtl, splitter

functional_only

none

abstains

merger, and2, or2, dro, clock_splitter

note

3 of 8 library cells carry a fitted model; the other 5 abstain rather than extrapolate.

spec_path_caveat

jtl, splitter answer margins ONLY when the numbers come from the spec-knobs head. On the netlist path there is no spec to read, the margins would come from the stimulus-artifact corpus, and all six quantities are refused. The distinction is the cell's LABELS, not the cell: see FINDINGS-2026-08-16-bias-margin-stimulus-artifact.md.

functional_head

model

gradient-boosted trees on spec knobs

test_auroc

dff

0.9761

jtl

0.9941

splitter

0.9993

escalation

split-conformal, per-cell qhat at alpha=0.10; answers only when the predictive set is an unambiguous singleton, otherwise defers to JoSIM

escalation_rate_test

dff

0.3232

jtl

0.0987

splitter

0.0667

n_cal

dff

1199

jtl

775

splitter

223

n_train

dff

2879

jtl

1861

splitter

535

caveat

AUROC is measured on held-out designs of the same fed corpus, not on designs outside its sampling box: dff at 0.9761; jtl at 0.9941; splitter at 0.9993.

margin_heads

model

gradient-boosted edge regressors; width is DERIVED as hi - lo so it cannot contradict the edges served beside it

metric

MAE vs the median-predictor baseline

improvement_over_median

null

improvement_basis

not quoted: the only measurement (2026-08-21, jtl and splitter, before the 4,800-row refit; dff not included) predates every head served now. Coverage below is read from the served bundle

conformal_coverage_on_test

dff

0.8712, 0.9105

jtl

0.881, 0.9071

splitter

0.9143, 0.9333

target

0.9

source

FINDINGS-2026-08-20-phase-b-corpus-regenerated.md, FINDINGS-2026-08-21-phase-c-heads-fit.md, FINDINGS-2026-08-21-phase-e-margins-served.md, FINDINGS-2026-08-26-qhat-fitted-quantile-is-not-an-operating-point.md

caveats

• the cell library is literature-reconstructed, not original foundry GDS
• a margin is a DISTRIBUTION under a named feeder, not a property of the cell type: width varies up to 2.17x across nine defensible reference feeders, and 2.7x (jtl) / 6.0x (splitter) across designs inside the operating band. Every served band names its feeder
• the raw-netlist path does NOT predict. The heads read spec knobs, so a bare .cir gets parse + structural DRC + the envelope check and then abstains
• escalation strength is only measurable where errors exist: jtl produces 3 errors on 720 test points and splitter zero, so their refusal curves are not yet evidence of anything. dff is the one cell where it can be measured

Refuses when

• outside the trained envelope -> abstained=true with null band factors, rather than a guess
• unknown cell name or knob -> HTTP 400 with the structured error

spice_characterize

One-call datasheet for a single RSFQ cell: margins, timing, yield and a surrogate cross-check.

Fidelity

analytic_transient_engine

Validation

reference

JoSIM. The quality gate (fluxus.neural.validate_spice_cells) requires per-JJ slip counts to match JoSIM exactly, and bias/Ic margin bands to agree within one sweep step, on JTL, splitter and comparator-DFF.

note

PASS/FAIL criteria are defined and the gate is runnable, but its results are not persisted as an artifact, so no measured figure is quoted here. Absence of a number here means no number was recorded -- not that the error is zero.

Refuses when

• unknown cell name or knob -> HTTP 400 with the structured error

sfq_netlist_margins

Bias, Ic and per-element (bias source, inductor, junction Ic) margins of any flat SFQ netlist under the caller's own functional test.

Fidelity

analytic_transient_engine

Validation

reference

the library margin walk (spice_cells.bias_margin) on the library JTL: the bisected high bias edge lands inside the walk's last 0.05 bracket (tests/test_sfq_netlist_margins.py). That is agreement with the same engine, not with JoSIM.

josim

confirm='josim' reruns the boundary points of the global knobs and of the critical element in JoSIM and serves each verdict beside the engine's; otherwise the edges are not checked against JoSIM for that netlist

note

no measured error figure is quoted: none is recorded for user netlists

Refuses when

• a netlist the engine cannot simulate (transmission lines, unsupported sources, symbolic values) -> HTTP 400 unsupported
• over a cap in envelope, an unknown port, junction or bias source, or a bad test definition -> HTTP 400 bad_request naming it
• the nominal run alone exceeds the budget -> HTTP 400 budget
• budget_s above envelope.budget_s.max, or a field out of range -> HTTP 422
• the netlist fails its own test at nominal -> HTTP 200 with refused_reason and no margins
• the run cap or the time budget is spent -> HTTP 200 with stopped and every knob not finished in not_swept; a knob cut off mid-walk is dropped, not served as an edge