There is a moment in every optimization where the transfer finally looks clean, and it is tempting to call the class done and move on. It is not done. A transfer that looks right can still be delivering the wrong volume, with a spread your eye cannot see, and the only way to know is to measure. Verification is the step that converts a tuned-looking class into a trusted one, and locking it in is what keeps that trust from evaporating the next time someone opens the editor. This is the end of the optimization loop, and skipping it wastes everything the earlier steps bought.
Measure delivered volume, do not infer it
Verification means putting real numbers on real deliveries. Two methods dominate, and they trade cost against convenience.
- Gravimetric testing weighs the delivered liquid on a calibrated analytical balance and converts mass to volume using the liquid's density at the measured temperature. It gives immediate, high-resolution feedback, which makes it the practitioner's default for closed-loop tuning, and it is the method most people reach for during optimization. Its main blind spot is stepped patterns like multi-dispense, which are awkward to weigh step by step.
- Photometric testing reads a dye, either a validated commercial kit or, for tuning purposes, ordinary food dye in a plate reader. Commercial dye systems are the accepted standard where regulatory acceptance matters; a food-dye-and-reader setup works comparably well for day-to-day optimization even though it lacks that formal standing.
Whichever you use, measure at your normal lab temperature and record it, because density, viscosity, and the delivered volume itself all move with temperature, and an unrecorded temperature turns a future discrepancy into a mystery.
Run enough replicates to see the spread
A single measurement tells you almost nothing, because you cannot distinguish a good class from a lucky shot. You need enough replicates to estimate both the average and the spread with some confidence. A practical working number is around ten replicates per volume, which is enough to make the standard deviation meaningful without turning verification into an afternoon; tighten that count upward when the assay tolerance is unforgiving. Run them across your actual working range, not just at one convenient volume, because a class that is perfect at 200 microliters can drift badly at 10.
Read precision before you touch trueness
Look at the spread first, then the average. This ordering is the same discipline that governs the whole optimization, and it is just as important at verification.
- Precision, the tightness of the cluster, comes from the mechanical parameters. If the standard deviation across your replicates is wider than the method allows, the class is not ready for a correction curve. Go back to the flow rate, settling time, and air handling, fix the spread, and re-measure. A correction curve applied to a wide spread just moves a blurry cluster onto the target; it is still blurry.
- Trueness, the distance from the average to the target, is what the correction curve fixes. Only once precision is acceptable do you adjust the curve so the average lands on target at each volume you measured.
Set the correction curve across the range
The correction curve maps the volume you request to the volume the instrument commands, cancelling steady over- or under-delivery. The single most common mistake is setting one point and assuming it holds everywhere. It does not, because the offset is rarely the same fraction at 50 microliters as at 300, so put a point wherever accuracy matters across your range and let the curve interpolate between them. For stepped workflows like multi-dispensing, correct against the actual volume remaining in the tip at each step, which keeps dropping, rather than the nominal aliquot size. After you set the curve, measure once more to confirm the average moved where you intended; a correction you do not re-verify is a guess with extra steps.
Know when to stop
Optimization has no natural finish line, so you supply one, and verification is where you check whether you have crossed it. The finish line is fitness for purpose, not perfection. Before you started, you should have decided what precision and trueness your assay actually needs; a normalization step and a qualitative screen do not demand the same tolerance. When the measured spread and offset sit inside that tolerance, across the volumes you run, the class is optimized. Stop. Chasing another half a percent past what the method needs is time the method will never pay back.
Lock it in so the tuning survives
An optimized class that lives only in one person's memory of what they changed is a class you will re-tune from scratch in six months. Locking it in has two parts.
- Document the result: write down the parameters, the tip, the labware, the temperature, the replicate count, and the measured precision and trueness at each volume. This verification record is what lets the next person trust the class without re-running your whole loop, and it is what lets you diagnose a future drift against a known baseline.
- Protect it from silent change: treat a calibration change as a real event, not a cosmetic edit. Anyone who later adjusts a parameter that affects delivered volume has invalidated the verification you just did, and the old measurements no longer describe the class in front of them. Make that consequence visible, so a tuned-but-unverified class cannot masquerade as a validated one.
That second point is the whole reason verification is worth doing. The evidence and the class have to travel together, because a class without its verification record is back to being a hypothesis, no matter how carefully it was once tuned.
A class you have not measured is a guess in a nicer font. Verify precision before trueness, correct across the range, then freeze the numbers and the evidence together.