Fluxus-S · SFQ logic
Deployed todayVerify SFQ logic
without the JoSIM queue.
Fluxus-S is the superconducting-digital half of Fluxus: margin, timing, yield, and design-rule verification for RSFQ single-flux-quantum cells — read from the netlist, measured against JoSIM, with a calibrated confidence bound on every prediction and a refusal where no head is fitted. It is our deployed product. 1,000 specs come back in 41 ms against roughly an hour of JoSIM — then read the baseline comparison below and decide where it actually helps you.
Delivery · API and CLI
A design loop, not an overnight queue.
Point the API or CLI at a cell and its spec knobs and you get a functional verdict and bias / Ic margin bands, each with a conformal interval, plus an explicit refusal wherever the design falls outside the box the heads were measured on. Batched, 1,000 design points come back in one call. Trained per process node on your reference simulator.
API + CLI · licensed per team
RSFQ circuit verification · MIT-LL SFQ5ee+
import requests
# one design point
r = requests.post(
f"{API}/api/sfq/predict",
json={"cell": "jtl", "params": {"bias_ua": 150.0}},
).json()
# a whole sweep in ONE call
b = requests.post(
f"{API}/api/sfq/predict/batch",
json={"cell": "jtl", "specs": specs},
).json()
# a number where a head is fitted, None where not
served = [x for x in b["results"]
if x["bias_low"] is not None]$ fluxus sfq-predict --cell jtl --cell splitter --cell dff
jtl bias 0.801-1.187 ±0.041 ✓
splitter bias 0.800-1.200 ±0.006 ✓
dff margins refused → no margin head fitted
jtl @ 3× bias — → outside envelope
and2 — → no model, abstains
─────────────────────────────────
$ fluxus sfq-batch --cell jtl --specs 1000.jsonl
1,000 specs 41 ms (JoSIM: ~59 min)
961 served · 39 escalated to JoSIM
41 ms
1,000 specs in one batched call, against ~59 min for the JoSIM margin runs it replaces
conformal
every served margin carries a band from a seeded calibration split
refuses
specs outside the measured envelope return no value
3 of 8 cells
jtl and splitter serve margins, dff serves its functional verdict only, the other five abstain
Capabilities · the whole SFQ surface
Every feature, with the number behind it.
Margin verification from netlists, with the uncertainty and refusal layer around it — each line carrying the metric that backs it.
Margins & verification
metric
Analytic margins
functional verdict + bias / Ic bands
Assembled-chain margin
shared-rail window, whole circuit
Three-engine validation
one operating point
ERSFQ / eSFQ bias
resistor vs zero-static sizing
Process portability
one netlist, retargeted
One-click datasheet
characterize a cell in one call
Uncertainty & throughput
metric
Conformal bands
per cell, from a seeded calibration split
Refusal, not a guess
out-of-envelope specs return no value
Chip-scale throughput
1,000 specs batched vs the JoSIM margin run
Figures are measured 2026-08-29 through the serving path, on the held-out 15% of a seeded shuffle of each cell's corpus — the same split the heads were fitted against, reproduced rather than re-drawn. 4,800 designs per cell for jtl and splitter, labelled by JoSIM under a current-triggered feeder; n = 720 test designs each, 180 for dff. The cell library targets SFQ5ee+ and is literature-reconstructed, not foundry GDS. The target is deterministic, so a high score is what a competent fit looks like, not a discovery — the number worth reading is the refusal behaviour below it.
The baseline, published
We published a baseline that beat us, so we shipped the baseline.
A gradient-boosted tree on tabular circuit features outscored our neural surrogate on every schema-derivable head. On 2026-08-29 the trees became the served model and the neural surrogate was deleted rather than kept as a fallback — measured, it had predicted functional=true on 100% of 40 designs and moved a margin edge by 0.019 across a 2× bias sweep. Below is what the shipped model measures now, against the only baseline that matters for a regressor: predicting the median. Lower MAE is better.
Two cells carry margin heads, so this is depth on jtl and splitter, not breadth across the library. dff serves its functional verdict and refuses its margins; five further cells abstain entirely. What the table establishes is that the heads are not fitting a constant — the labels carry signal and the margin over the median predictor is real.
Across the eight margin heads the improvement over the median runs 4.7× to 42×, and every conformal band covers between 0.869 and 0.916 against a 0.90 target. Read that as a competent fit, not a discovery: JoSIM is deterministic and the spec fully determines the netlist, so there is no noise floor to beat. It reproduces the oracle. Nothing here is calibrated against fabricated silicon.
So the model is the boring part. What Fluxus adds is the layer around it: a conformal band on every prediction from a seeded calibration split, a refusal when a spec falls outside the measured envelope, and a per-quantity refusal where no head was fitted at all. That layer is measurable, and it was measured — every one of the three errors jtl makes on 720 held-out designs lands inside the set it declined to answer, and dff catches two of its four. A number you cannot audit is worth less than a blank you can act on, which is why the capability manifest your agent reads names the classical baseline first.
Process nodes
Validated where it counts.
Fluxus-S targets MIT-LL SFQ5ee+ — measured against JoSIM on a literature-reconstructed cell library for that node, not on foundry GDS — and retargets to other fab nodes from one netlist. Per-process calibration is a fixed-scope engagement: we train on your reference simulator, on your library, and hand you the endpoint with its validity report.
Building SFQ control for qubits?
Fluxus-S and Fluxus-Q together enable co-design — simulating SFQ control logic against qubit Hamiltonians for flux crosstalk, timing, and heat load at 10 mK. Own both, or start with the half you need now.
Designing RSFQ logic?
Let's talk.
If your design loop is bottlenecked on JoSIM margin runs, we'd like to understand it — and train Fluxus-S on your process.
Get in touch