The 10 µF Capacitor That Measures 4 µF

Put 3.3 V across a small ceramic capacitor marked 10 µF and you may be running your circuit on 4 µF or less. Nothing is broken. The part is in spec, the reel is genuine, and the loss is printed in the manufacturer's own curves, just not on the one-line entry your CAD library shows.
The number on the label is measured the way the makers specify it: a small AC test signal near one volt, at 1 kHz or 120 Hz, with zero DC across the terminals. Your circuit never operates the part that way. A decoupling cap sits at the rail voltage all day, and for Class II ceramics that steady voltage is what eats the capacitance.
Where the loss comes from
X5R and X7R parts are built on barium titanate, a ferroelectric ceramic. Its huge permittivity comes from electric domains inside the grains: small regions of aligned polarization whose walls shuffle back and forth when a signal wiggles the field. That shuffling is the capacitance.
A steady DC field clamps the shuffle. Domains line up with the bias and stop responding to the small signal riding on top, so the effective permittivity falls, and the capacitance falls with it. The effect scales with field strength inside the dielectric, volts per micron, not with the voltage rating printed on the reel. A 6.3 V part biased right at its rating sits deep in the bend of that curve; the same part held at 1.8 V has barely left the flat.
That distinction decides everything that follows. Two 10 µF, 6.3 V X5R capacitors from the same maker, one in an 0805 body and one in an 0402, see different internal fields at the same 3.3 V rail, because the smaller part packs its capacitance into thinner layers. The densest modern parts run dielectric layers around a micron or under. Thinner layer, more volts per micron, deeper loss.
Class I dielectrics, C0G and NP0, are paraelectric. No domains, so no bias loss to speak of, and no aging either. You pay for that in capacitance per cubic millimeter, which is why nobody decouples a processor with C0G.
The case-size trap
The place this bites hardest is a substitution made under schedule pressure, and shortages produce those weekly. Say the BOM calls for a 10 µF, 6.3 V X5R in 0603 on the output of a small buck converter running at 3.3 V. The 0603 goes on allocation, an 0402 with the same three headline numbers shows up in stock, and on paper it is a drop-in: same capacitance, same rated voltage, same temperature class, even the same maker. Then boards built with the 0402 show higher output ripple, and a converter that was stable starts ringing under load steps. Pull the bias curves and the story is plain. At 3.3 V the 0603 might hold on to something like half of its nominal value while the 0402 keeps a third or a quarter, figures that shift by series and by maker, which is exactly why you check the curve for the precise part number instead of trusting the family. The output capacitance sets both the ripple voltage and the position of the LC corner the control loop was compensated around; cut it in half again and the crossover moves, phase margin shrinks, and a marginal design tips over. None of this shows up at incoming inspection, because an LCR meter on the bench applies the same zero-bias condition the datasheet used, reads 9-point-something microfarads, and passes the lot. The board misbehaves in circuit while every individual part measures fine, which is the kind of fault that burns a week of two engineers' time. The defect was created the moment someone approved an alternate based on three matching numbers in a line-item table.
Aging runs the other way
Bias loss is instant and recovers the moment the voltage is removed. Class II parts carry a second, slower drift on top of it: ferroelectric aging. Capacitance decays by a few percent for every tenfold stretch of time after the ceramic last cooled down through its Curie point, which for barium titanate sits near 125 °C. Makers commonly reference the nominal value to a set time after that event, on the order of a thousand hours, so a reel that sat in a warehouse for two years can measure a few percent low and still be in spec.
Reflow resets it.
A trip over the Curie point during soldering de-ages the ceramic, capacitance pops back up, and the slow decay starts again from hour zero. So an old date code on a Class II MLCC is rarely an electrical problem in itself; the part you solder behaves like a fresh one. Solderability of the terminations is a separate question, and it belongs to storage conditions, not to the dielectric.
What to do at design time, and what to ask before you buy
At design time, pull curves, not classes. Murata publishes bias data in SimSurfing, TDK in SEAT, KEMET in K-SIM, and the other majors have equivalents. Look up the exact part number at your operating voltage and write the derated value, not the nominal one, into your margin calculation.
A rough rule that has aged well: keep the DC working voltage under half the rated voltage, and treat anything tighter than that as a flag for a curve check. Going up one case size at the same capacitance and rating buys back a surprising amount, since the bigger body uses thicker layers.
The catch: bigger cases spend board area you may not have.
At purchase time, qualify alternates by curve, not by line item. Matching capacitance, voltage, and temperature class across two case sizes, or across two makers, tells you nothing about behavior at your rail. When sourcing through distribution, give the full manufacturer part number and ask for it back on the paperwork; a broad-line independent distributor such as In Fortune Electronics can quote against the exact ordering code, and the exact code is the only level at which a bias curve means anything.
And if a board that passed every bench check develops ripple or instability after a passive substitution, measure the capacitor in circuit, under bias, before blaming the silicon. The 10 µF that reads 4 is sitting right there.