Bias sweep · jtl · captured
Move the bias until it stops answering
41 design points of a JTL, from 0.35× to 2.2× its nominal 150 µA bias, each one a real call captured from the serving path. Near nominal you get a band. Toward either edge the answer degrades the way it should: a failing verdict, then an abstention. Nothing is extrapolated.
0.804–1.188×
bias margin window
±0.041
conformal half-width
1.000
P(functional)
Functional, with both margin edges served and a conformal band on each.
Across the sweep: 10 answered · 6 not functional · 25 abstained
One batched call · captured
1,000 jtl specs in 64.8 ms
The JoSIM margin characterization this replaces runs about 3.56 s per functional design, so the same 1,000 specs take roughly 59 minutes. Every spec still gets its own envelope gate and its own conformal decision; batching changes the cost, not the answer.
1,000
design points
64.8 ms
one call
951
answered
49
escalated to JoSIM
Served bias-window width across the batch (median 0.422)
≈ 54,938× faster per spec than the JoSIM run it replaces. Reproduces the JoSIM oracle, not silicon.