Pre-loaded design points
5 results, and 2 of them are refusals
Fluxus-S carries a fitted head for 3 of 8 library cells. Reproducing a deterministic oracle accurately is table stakes; knowing which designs it may not answer for is the product. Each of these is a real call against the serving path, captured on 2026-09-02 and replayed here verbatim.
jtl
jtl
Josephson transmission line. Both margin edges are served with a conformal interval fitted on a held-out calibration split.
splitter
splitter
Fluxon splitter. Same heads, a visibly tighter band -- the band is fitted per quantity, not assumed.
dff
comp_dff
Clocked-comparator DFF, the library's only SEQUENTIAL cell. Margins shipped 2026-09-02 on a purpose-built 4,800-design corpus and measure 3.7x-4.2x over a median predictor -- below jtl's 4.7x-8.8x and splitter's 18x-42x, and quoted per cell rather than averaged. Its band escalates far more (32% of designs) and in exchange the answers it does give carry 0.8% risk against 7.2% if it answered everything.
jtl_out_of_envelope
jtl
The same cell at 3x its nominal bias. Outside the box the heads were measured on, so it returns no number and says which field, what value, and what range was allowed.
and2
and2
An AND2. Five of the eight library cells have no fitted model, and they abstain rather than extrapolate from the three that do.
Captured from the Fluxus-S serving bundle through the same service functions the API serves, so nothing on this page is authored by hand. Reproduces the JoSIM oracle, not silicon.