A cloud of plausible next blocks, redrawn several times a second and collapsed onto each block that actually arrives. Scores how far outside its own prediction the chain lands.
Predicted load, as forty outcomes
The same prediction the cloud was showing when the last block landed, projected onto the transaction count and held still. Each dot is one outcome in forty, so counting the dots to the left of the marker gives the chance the prediction assigned to a block that small. This is the reading that survives reduced motion, where the cloud is drawn once per block rather than continuously.
Nothing to plot yet. A prediction has to be on screen when a block lands before there is anything to compare, which takes 12 blocks of warm-up and then one more.
Where the surprise scores landed
The Calibration figure above is the average of these bars, and an average hides the shape. A predictor that is far too confident half the time and far too cautious the other half averages 50% just as a well-behaved one does. Flat bars mean the block really does fall anywhere in the cloud with equal chance, which is what calibrated means; bars piled at the right mean the chain keeps landing outside a prediction claiming to be tighter than it is. The dashed line is where every bar would sit if the prediction were perfect.
The ± beside the calibration figure is the same caution applied to the figure itself. It is the 95% interval around that average, so a session of twenty blocks quotes a range wide enough to cover most of the scale — which is the honest reading, because twenty blocks cannot tell 42% from 58%. Only once the interval clears 50% entirely is the predictor measurably off.
- 0% — landed dead centre
- 0 blocks scored
- 100% — landed outside every draw
The cube’s three axes describe a block. Every few hundred milliseconds the cloud is redrawn as a completely new set of plausible next blocks, each one a recent block picked at random and nudged — so the cloud is the shape recent behaviour has actually had, with no distribution assumed. When a block lands, the cloud collapses onto it and the trail gains a point. The scale is taken from the last 96 blocks, so the whole trail is redrawn as the chain’s behaviour shifts.
- AcrossLoad — how many transactions the block carried, on a log scale330
- UpComplexity — contract calls as a share of all transactions36%
- DepthCost — energy burned by the block in TRX, on a log scale78
Energy is read at block level, because core reports it per block rather than per transaction. With reduced motion switched on the cloud is drawn once per block instead of continuously, which still shows the spread but not the shifting that gives the display its name. The dotplot above is unaffected by that preference, and it is the better reading either way for a viewer who wants a number rather than an impression.