Sensitivity

Which values move each goal, and by how much — ranked at the design as it stands, before you tune, optimize or estimate a yield.

What does the design depend on?

The Sensitivity panel (View ▸ Panels ▸ Sensitivity, the tab after the Optimizer, where the goals are set) answers that question for the focused schematic, at its values as they stand. Press ▶ and every value is moved a small step up and a small step down on its own, with everything else held where it is. Each goal is then given a ranked list of the values that move it, with each one's share of the total.

That is 2k + 1 simulations for k values — the design once, then each value twice — and nothing is drawn at random, so the same design always gives the same ranking. The cost is shown under the settings before you run: 15 factor(s) (7 at ±0.1 dB, 7 at ±1°, 1 at ±σ), 31 evaluations.

The panel is the Yield panel's view with one mode. It edits the same schematic: a tolerance added here is in the Yield panel at once, and ⌘Z undoes it from either. Nothing runs until you press ▶, and a run never changes the schematic's values; its result is written beside the schematic as <name>.sens.npy.

Where it fits: Sensitivity, then Tuning and the Optimizer, then Yield

The panels sit in the order a design is usually worked:

1. Sensitivity — find the few values that matter.
2. Tuning and the Optimizer — move those few, by hand or by search, until the goals are met.
3. Yield — give the values that matter a tolerance and count how many builds pass.

The Optimizer's Sensitivity button opens this panel on the same schematic when no optimization has run. After a run, the same button ranks the variables around the optimizer's best point instead.

What is stepped, and by how much

A port's reference impedance and a source's tone frequency are left out of All; each run says which values it left out and why. Each setting is saved with the schematic as the sensitivity line, and only a value that differs from its default is written.

S-parameter blocks

Every S-parameter of a Touchstone block is a value the ranking can move: its magnitude in dB and its phase in degrees, offset the same at every frequency. Before any run the panel lists each block under S-parameter blocks — 2 ports · tied · 6 factors — so a block you are not going to change can be left out before its simulations are spent. Untick Include and the cost under the settings drops by two simulations per factor.

Untick S-parameters and every block is left out at once. Each block's Include box stays as you set it, greyed, and ticking S-parameters again brings those choices back.

A block whose file is reciprocal — S21 equal to S12, as for any passive part — offers one value for each pair: moving S21 moves S12 with it. A block that is not, such as an amplifier, offers every entry on its own. Tie reciprocal pairs, on the matrix tile, changes that.

Reading the ranking

Choose a goal — under Specs, or in the picker above the list — and the list shows what moves it, largest first:

Each goal row under Specs shows m₀, its normalized margin at the design as it stands (positive means met), and σ, the spread the ranking predicts (√ΣΔ²). When every value was stepped by its own tolerance, a first-order pass estimate follows, labelled estimated. The yield itself is what the Yield panel counts.

An S-parameter block you left out is the last row, greyed, with excluded for its basis and no share: it was not found insensitive, it was not stepped. Tick its box to include it in the next run. A run does not repeat on its own, so until you press ▶ again the status line says the ranking is out of date.

The basis column

A ranking can mix two kinds of step, and the basis says which each row had:

• σ — the value moved by its own tolerance. Its share says how much of the goal's spread in production comes from that part.
• %, dB, ° — the value moved by a fixed small step. Its share says how strongly the goal responds near here: what a designer could change, not what production will do.

When any row is not σ, there is no pass estimate, because σ is then not a production spread.

Entry, Port and Instance

The same run, grouped three ways — the toggle regroups the list without simulating again:

Nonlinear, one-sided and failed

A glyph at the end of a row flags a value whose Δ needs care; its tooltip gives the reason.

The S-parameter matrix tile

Select an S-parameter block, or any of its rows, and a tile opens under the list: one cell per entry, coloured by that entry's share of the selected goal, on one scale from 0 to the tile's largest share. The selected entry is outlined, and at the Port level so are the selected port's row and column. A tied block colours S21 and S12 as one. Each cell's tooltip gives its share.

Two settings ride on the tile:

• Include — the same setting as the block's row under S-parameter blocks.
• Tie reciprocal pairs (S21 = S12) — whether one value moves both entries of a pair. Its tooltip shows what the file decides on its own (auto: worst |S(i,j) − S(j,i)| = 0.0021). Choosing what the file decides returns the block to automatic; choosing the other is saved on the block as Reciprocal=tied or untied. Tying ends the entries above the diagonal, so when a tolerance names one the panel lists those lines and asks first; the change and their removal are one undo step.

Acting on a row

Right-click a row:

• Add tolerance gives a value that was stepped by a percent, dB or degrees a tolerance, the same as the Yield panel's +. The next run steps it by σ.
• Send to Tuning adds the value to the Tuning panel.
• Use in Optimizer ticks Opt with the default range.

S-parameter offsets are one number at every frequency, so they are tolerances only: Send to Tuning and Use in Optimizer are greyed on their rows.

Run Yield on these opens the Yield panel on the same schematic in Monte Carlo mode. When some of the ranked values have no tolerance, its status line says how many: a Monte Carlo holds them still.

The display

The chart button writes <name>.sens.cdd beside the schematic and opens it: one Pareto chart per goal at the level the list shows, the shares as bars, largest first, with their running total as a line. The Port level has no running total, because its shares overlap.

The example

File ▸ New ▸ Examples ▸ Yield, Corners and Centering includes SnpSensitivity: an amplifier's Touchstone block followed by a spiral inductor's, with a toleranced shunt capacitor and two specs. The spiral's file is reciprocal and the amplifier's is not, so the two tiles show a tied and an untied block. Its README says what to click.

From the command line, and from an agent

circuitrf yield sens <schematic> runs the same ranking and writes the same .sens.npy; --by entry, port or instance prints a level. An agent runs it through MCP as run analysis=sensitivity. See the command line and AI agents.