Fluxus-S

Served

RSFQ cell margins

Whether an RSFQ library cell works at a given set of design knobs, and — only beside a functional verdict — its bias and critical-current (Ic) margin bands, each with a conformal interval.

For: RSFQ / SFQ logic designers whose loop is bottlenecked on JoSIM margin runs.

427×

faster than the JoSIM margin run

8.8 ms per call vs 3.78 s · single design

Status. Served for 3 of 8 library cells; the other five abstain.

Engine. Data-trained surrogate of JoSIM transient simulation (gradient-boosted trees — the baseline that beat our neural model, now the served model).

Reference. JoSIM transient simulation of a literature-reconstructed RSFQ cell library targeting the MIT-LL SFQ5ee+ process. Not foundry GDS, and nothing is fab-validated.

API

Endpoints

POST/v1/fluxus/sfq/predict

one design → verdict and margins

POST/v1/fluxus/sfq/predict/batch

up to 5,000 designs in one call

GET/v1/fluxus/sfq/schema

cells, knobs, units and ranges

Base URL https://api.cogitan.ai, bearer key from an approved account. Read GET /v1/fluxus/capabilities first: it states the same fidelity, validation, envelope and refusal conditions as this page. Access model.

Inputs

What you send

cell

jtl | splitter | dff (served)

jtl · bias_ua

90 – 180 · µA

jtl · l_series_ph

1.5 – 4 · pH

jtl · area

1.6 – 2.8 · ×

splitter · bias_in_ua, bias_br_ua

120 – 180 · µA

splitter · l_branch_ph

1.5 – 3.5 · pH

dff · bias_in_ua, bias_q_ua

100 – 180 · µA

dff · bias_f_ua

40 – 130 · µA

dff · bias_esc_ua

110 – 200 · µA

dff · lq_ph

12 – 32 · pH

Outputs

What comes back

functional

verdict and probability

bias_low / bias_high

bias margin band, fraction of nominal, with conformal interval

Ic low / Ic high

critical-current margin band, with conformal interval

escalation

a flag that hands the design to JoSIM instead of answering

Validation

Measured against what, and how

Reference. JoSIM transient simulation of a literature-reconstructed RSFQ cell library targeting the MIT-LL SFQ5ee+ process. Not foundry GDS, and nothing is fab-validated.

Sample. 4,800 simulated designs per cell.

Split. Seeded random 60/25/15 train / calibrate / test split. Not grouped: the target is deterministic, so a high score is what a competent fit looks like, not a discovery.

JoSIM — Josephson junction superconductive SPICE circuit simulator

dff — designs escalated to JoSIM

~32%

dff — error rate on answered designs

0.8%

48 of 52 errors caught by escalation

jtl — errors caught

3 of 3

every error was in the declined set

Held-out test split, measured through the serving path (2026-09-02)

Cell	n test	Accuracy	Majority baseline	AUROC	Margin MAE vs median	Conformal coverage (90% target)
jtl	720	0.9958	0.5847	1.0	4.67–8.76× better	0.869–0.897
splitter	720	1.0	0.7972	1.0	17.89–42.19× better	0.882–0.916
dff	721	0.9279	0.6519	0.9761	3.69–4.24× better	0.871–0.911

dff coverage is from the served bundle (2026-10-02), which supersedes the benchmark file.

Speed

What it costs per call

One design point, warm

427× faster

JoSIM margin run

3.78 s

Fluxus

8.8 ms

Single call

8.8 ms

vs 3.78 s for the JoSIM margin run it replaces (427×)

Batched

~0.043 ms / spec

1,000 specs in one call

Envelope

Where the answer is trusted

cells

jtl, splitter, dff — and2, or2, merger, dro and clock_splitter abstain

knobs

per-cell ranges above; GET /v1/fluxus/sfq/schema is authoritative

batch

≤ 5,000 specs per call — larger requests are truncated and say so

Refusals

When you get no number

Outside the knob envelope: abstained = true with null margin bands, never a guess.
Margins are served only beside a functional verdict. A design predicted not to work gets no margin, with the reason stated — JoSIM defines no margin for a design that does not work.
Unknown cell or knob: HTTP 400 with a structured error.
The five abstaining cells return no model rather than an extrapolation.

Known limits

What it does not do

Coverage is 3 of 8 library cells. The rest abstain.
A margin is a property of the cell under a stated reference drive, not of the cell type: margin width varies up to 2.17× across defensible feeder / drive conditions.
It reproduces the JoSIM oracle, not silicon. The cell library is literature-reconstructed for the SFQ5ee+ node, not foundry-validated.
The split is a seeded random split, not grouped. That is defensible here because the target is deterministic, but it is not an out-of-distribution test.

Where something else wins

A gradient-boosted tree on the spec knobs beat our neural surrogate on every schema-derivable head. The trees are what is served; the neural surrogate was deleted rather than kept as a fallback.

Every figure on this page compares against a simulator or a published device measurement. None of it is a prediction of your fabricated hardware. For one certification-grade answer, run the full reference; this is for screening many designs cheaply and spending solver time on the survivors.