The answer is a difference — and differences amplify noise

Differential 2D measurement recovers the sheet as composite minus bare support. Each dot below is one simulated differential pair at one value of the sheet’s share of the composite’s transport; every measurement carries the same per-measurement noise. Where the sheet carries most of the signal (right), the recovered values hug the true line; where it carries a sliver (left), the identical instrument scatters wildly — the exact error law for this model is σ/share·√(1+(1−share)²), which approaches √2·σ/share for small shares. The shaded funnel is the ±1σ prediction of exactly that law.

Preset:
one differential pair each±1σ funnel: σ/share·√(1+(1−share)²)true sheet value

Model: sheet value recovered as the difference of two independent measurements — composite (conductance 1/share) and bare support (1/share − 1) — each with independent Gaussian relative noise σ. The recovered value’s exact relative SD is then σ/share·√(1+(1−share)²): at full share it reduces to σ, and for small shares it approaches the √2·σ/share limit. This is the equal-variance, independent-error teaching model; real campaigns add support-reproducibility, coating and covariance terms (paired designs on the same membrane can be positively correlated, reducing part of the variance while transfer processing adds its own). Axes normalized; schematic teaching tool, not an instrument.