If you ask most IB Physics students what makes a strong Internal Assessment, they'll talk about the experiment: an interesting research question, a clever setup, maybe a pendulum with a twist or a homemade diffraction rig. Almost nobody's first instinct is "my uncertainty propagation." And yet that's usually where marks quickly disappear. I learned this the hard way on my own IA draft, where the experiment itself was solid but the error analysis was let's say questionable. Here's what I'd tell anyone starting theirs now.
Why examiners care so much about this
The IA is marked against five criteria, and two of them — Analysis and Evaluation — are almost entirely about how rigorously you've treated uncertainty. It's not enough to record data and draw a best-fit line. Examiners are checking whether you understand why your result differs from a theoretical value, and whether your stated uncertainty is actually consistent with your data, rather than a token "±5%" pasted at the end because a percentage looked appropriate.
A genuinely good IA treats uncertainty as part of the investigation, not a formality at the end of it. That distinction is what separates a 6-mark Analysis score from a 3-mark one.
The mistakes that show up most often
Uncertainty that doesn't match the equipment. If you measured a length with a ruler, your reading uncertainty is realistically ±0.5 mm (half the smallest division), not an arbitrary ±1%. Examiners notice when stated uncertainties don't trace back to an actual instrument limitation.
Propagating errors incorrectly, or not at all. If your final quantity is calculated from several measured variables — say, a gradient used to extract a constant — the uncertainty on that constant has to be propagated through the relevant relationship (usually via partial fractional uncertainties, or from the max/min gradient method on a graph). Skipping this step and just eyeballing an error bar is one of the most common ways marks are lost.
Confusing random and systematic error. Repeating a measurement five times only addresses random error; it does nothing for a systematic offset like a miscalibrated sensor or parallax error introduced by your own setup. A strong evaluation names the type of error present and explains what it would take to reduce it, not just a generic "human error" line, which examiners actively penalize as vague.
Not connecting the numbers back to the conclusion. The most common structural mistake: a student calculates a percentage difference from the accepted value, states it, and moves on without asking whether that difference is actually explained by the propagated uncertainty. If your result is 15% off but your uncertainty is only ±2%, that's a real discrepancy that needs a physical explanation. If your uncertainty is ±18%, the discrepancy is fully accounted for by experimental limitation, and you should say so explicitly.
A workflow that actually works
Decide your uncertainties before you take data, not after. Know what your instruments can and can't resolve, and write down your reading and calibration uncertainties before the first trial. It forces you to think about precision as part of the design, which usually improves the design itself.
Repeat every measurement enough times to estimate random error properly. Three trials is the bare minimum; five gives you a standard deviation that actually means something.
Propagate uncertainty through every calculated step, not just the final answer. If you compute intermediate quantities, carry their uncertainty forward.
Use max/min gradient lines on graphs, not just a single line of best fit, to get a defensible uncertainty on any gradient-derived quantity.
In your evaluation, separate "what limited my precision" from "what limited my accuracy." These are different questions with different fixes, and conflating them is one of the fastest ways to lose Evaluation marks.
Suggest an improvement that's actually specific. "Use more accurate equipment" is not a real suggestion. "Replace the manual stopwatch timing with a photogate sensor to remove reaction-time error, which was the dominant source of uncertainty in Trial 2" is the level of specificity examiners want to see.
The bigger point
Error analysis isn't a box to check after the interesting part of your IA is done.The best Physics IAs I've read don't have the flashiest experiments; they have the most honest ones, where the student clearly understood exactly why their number came out the way it did, down to the millimeter or the millisecond. That's the skill examiners are actually looking for, and it's learnable well before you touch your final write-up.