A liquid class is off. Fine, but off how? That is worth pinning down before you touch a single parameter, because accuracy and precision fail for different reasons and rarely at the same moment. One class lands on 47 microliters every time you ask for 50. Another averages a flawless 50 and sprays anywhere from 44 to 56. Same complaint, two different problems, and the afternoon people lose is almost always the one spent nudging a flow rate to fix a bias that a correction curve would have cleared up in five minutes.
Two different questions
Accuracy is about the average. Does the mean delivered volume land on target, or does something shove every dispense the same way, a hair short or a hair long each time? That shove is bias. It is systematic, meaning it keeps its direction and roughly its size from one dispense to the next. Precision is a different question and does not care about the target at all. It asks how tightly the individual dispenses bunch up. You can be tight and wrong, clustered neatly around 47 when you wanted 50. You can also be right on average and hopeless shot to shot, centered on 50 with replicates flung across a six-microliter spread. Most struggling classes carry a little of each. What matters is that you pull them apart and diagnose them one at a time.
How to see each in the data
You cannot eyeball this one. Weigh a set of replicate dispenses, turn the masses into volumes, and let two summary numbers do the talking. The mean gives you accuracy: subtract the target and what is left is the bias, in microliters or as a percentage. The spread gives you precision, almost always quoted as a coefficient of variation, which is just the standard deviation over the mean, written as a percent. Report both. A lone average will cheerfully hide a class that scatters by ten percent, and a lone CV will just as cheerfully hide one that groups tightly around the wrong number.
Which knobs move which
Once you know which failure you are looking at, the fix is less mysterious than it seems. A handful of settings move bias. A different handful move scatter. Very few move both.
- Bias, the wrong mean: go to the correction curve first. It exists to cancel a steady offset, mapping the volume you request onto the volume the instrument actually commands. Over-aspiration helps as well, drawing a touch of excess to make up for liquid the tip holds back. Both move the mean and tend to leave the scatter alone.
- Scatter, the loose cluster: here you are trying to make every dispense look like the one before it. Slow the flow rate. Give the liquid a moment to settle before the tip lifts out. Size the air gap properly, switch on level tracking, and pre-wet the tip so the first shot is not an outlier that drags your CV up by itself.
- A few settings pull in both directions, so change one thing per run and reweigh. It is slower. It is also the only way you ever learn which knob actually moved the number.
Why it matters downstream
None of this is bookkeeping. A biased class quietly poisons the answer itself: every concentration comes out shifted, a standard curve built on it tilts, and the error tucks itself neatly inside a result that looks spotless. Scatter does something else. It leaves the answer where it is and makes it harder to believe, pumping up the variance until a genuine effect sinks into the noise between replicates. So the two failures bill you in different currencies. One hands you a wrong number you trust. The other hands you a right number you cannot defend. Work out which one a class has and you know, before a single real sample goes on the deck, what your data will and will not be allowed to claim.
Diagnose before you tune. A class that misses the target and a class that cannot repeat itself look identical in a single average, so measure the spread before you touch a thing.