A model can be wrong in two completely different ways — and "more accurate" and "more precise" are not the same fix. Four target boards fill themselves as you scroll, two sliders let you take the gun, and a quick game asks you to call each pattern. Then the part most people miss: why a method that is precise but wrong can still tell you exactly which way things are moving — and why "just make it both" is sometimes not on the menu.
The bullseye is the truth — the value a model is actually trying to hit, on repeated tries with slightly different training data. Ten shots land on each board on their own. Where they cluster relative to the centre is accuracy; how tight they cluster with each other is precision.
Drag the sliders and the shots move as you drag — accuracy walks the cluster in toward the bullseye, precision tightens or loosens it. Watch how the same "good model" feeling can come from very different combinations, and how low precision alone can hide a model that's accurate on average but unreliable on any single prediction.
Start with the only thing that matters: which of the two properties does this pattern have? Once those two calls come easily, the jargon underneath them is just a lookup.
An unlabelled pattern. Make both calls independently: does it land on the truth, and does it repeat itself?
You've got the two axes. Now try putting a name on the combination ↓
Same idea, harder ask. Combine the two calls into the label a modelling team would actually use.
Every one of these four patterns has a direct counterpart in how a model learns from data — and the fix for one is often the wrong fix for another.
Here is the part most people skip. Twelve months of a real quality score, measured two ways. Instrument A is precise but biased — it reads about nine points high, every single time. Instrument B is accurate but noisy — right on average, wrong on any given month. Read the level and A looks useless. Read the change and A is nearly perfect, while B falls apart.
Quality score, absolute level
A sits well above the truth all year. It is plainly wrong.
Same data, each series minus its own month 1
A's line lands on the truth. B's is a mess. The ranking flips.
Level error = how far the instrument's average sits from the truth's average. Trend error = how wrong its reported change-since-January is, averaged over the year. Both in score points.
| Month | True | A reads | B reads | True Δ | A Δ | B Δ |
|---|
A constant bias subtracts out of a difference. Random noise does not. If your method is precise but off, you still cannot say what the number is — but you can say exactly which way it is moving and by how much. That is most of what a dashboard is actually for.
The inverse is the trap nobody expects: an accurate-but-noisy method is trustworthy as a level and useless as a trend. Two noisy readings a month apart tell you almost nothing about the month — you have to average enough of them first, and averaging costs time you may not have.
The same split shows up in model evaluation, in instruments, and in the metrics on your board deck. In every row, the third column is the move that buys you the property you're missing.
| The thingMeasured how | Missing accuracyPrecise, not accurate | Missing precisionAccurate, not precise | The moveHow you get both |
|---|---|---|---|
| AILLM-judge rubric | A judge model that grades 0.8 points harsh on everything. The absolute scores are wrong, but rankings and week-over-week deltas hold perfectly. | Three human raters whose average lands on the true score but who disagree with each other by ±1.5. Any single rating is a coin flip. | Anchor the judge on a small hand-graded gold set — that buys accuracy. Then average five samples per item — that buys precision. Two different fixes. |
| AIDemand forecast | A model that comes in 12% under every quarter because it never learned Q4 seasonality. Reliable, and reliably low. | An unbiased forecast with a ±30% interval. Right in expectation, unplannable in practice. | Add the missing seasonal term to kill the bias, then ensemble across models to shrink the interval. |
| InstrumentBathroom scale | Reads three pounds heavy every time. Your weight is wrong — and your week-over-week change is exactly right. | A wobbly scale that averages correctly but swings ±4 lb depending on where you stand. | Put a known 20 lb weight on the precise one. A single measured offset corrects every future reading, forever. |
| InstrumentGPS position | A receiver set to the wrong map datum: every fix is 2 m off in the same direction, so every distance between fixes is exact. | Consumer GPS — correct averaged over a few minutes, but any single fix is ±5 m. | Differential GPS: a base station at a surveyed position broadcasts the correction. Precision from the receiver, accuracy borrowed from a known point. |
| BusinessNPS / survey | A panel that consistently over-samples happy customers. The absolute score is inflated; the quarter-on-quarter move is real. | A methodologically clean survey with n=80 and a ±8 point margin. Your "6 point improvement" is noise. | Re-weight the panel against a census to remove the tilt, then raise n to tighten the interval. Sampling frame first, sample size second. |
| BusinessCycle time | A clock that always starts 40 minutes late because the ticket is created after the work does. Every duration is short by the same amount. | Hand-logged times with the right average and enormous variance, because people write them down whenever they remember. | Fix the trigger event so the clock starts at the real start, then instrument it so no human has to log anything. |
Precision comes from inside the system — more replicates, a tighter process, averaging, a longer sampling window, a stronger regulariser. You can always spend your way to a tighter cluster.
Accuracy can only come from outside it — a reference standard, a gold-label set, a known weight, a surveyed base station, ground truth. No amount of repeating a measurement will ever tell you it is centred.
Which sets the order of operations: get it precise first, then calibrate. A precise instrument needs one correction, measured once. A noisy one needs a correction you cannot even see, because its own scatter is hiding the offset you're trying to find.
"Just make it accurate and precise" is sometimes not a choice anyone is declining to make. Six structural reasons the top-right corner of the board is closed — and what you do instead when it is.
Every knob that cuts variance — regularisation, smoothing, a simpler model, a shorter window — pays for it by introducing bias. That's the mechanism, not a bug. You're picking a point on a curve, and minimum total error is almost never at zero bias.
Precision is measurable from repeats alone. Accuracy needs a true value from outside. For "is this summary any good?" or "will this customer churn in 2028?" there is no certified weight to put on the scale — so you can prove consistency and never prove correctness.
When expert annotators disagree 12% of the time, the truth has a radius. Accuracy is capped at that irreducible error, and every point you claw back past it is fitting your labellers rather than reality.
A scale that can weigh a truck cannot resolve a gram. Widen the range and the steps get coarser. And measuring disturbs what is measured — the probe warms the sample, the survey changes the answer, the metric becomes a target and stops measuring the thing.
Halving the noise takes four times the samples — expensive, but it's just money and compute. Removing the bias takes ground truth somebody has to produce, which is the slow, political, frequently-blocked part. Budgets and calendars pick one.
Sensors drift, distributions shift, and the population you calibrated on isn't this quarter's. A model that was accurate in January is biased by June unless someone re-anchors it — which is exactly why precision-plus-trend is often the more durable thing to run your decisions on.
Teams often chase a single accuracy number and miss that the real problem is variance, or chase tighter predictions and miss that they're now consistently wrong — or throw out a biased-but-precise measure that was the best trend signal they had. That's the kind of evaluation question we help teams get right before they ship.
Talk to us about your AI roadmap