Reading results
What the histogram is showing you, why its axis stops short of the extremes, and which numbers are measured over everything.
A simulated node reports a range, not a figure.
The histogram shows the middle 98%
The chart is binned over the central 98% of the samples — from the 1st percentile to the 99th — across 50 bars.
Without trimming, one draw in ten thousand landing far out would set the axis and squash everything else into a single bar.
What happened to the rest
Samples outside that range are counted separately, as below and above. They are not folded into the edge bars, because doing that would invent a spike the distribution does not have.
The chart names them at the edge — how many there were and how far they actually reached — but only when they lie more than one bar beyond the axis. Every continuous distribution leaves its outer 1% off each end, so saying so every time would be noise.
The axis labels tell you which case you are in
A second row under the chart names what the axis ends mean:
| Labels | Means |
|---|---|
| P1 / P99 | The range was trimmed — there are samples beyond the axis |
| Min / Max | The full range fits; nothing was trimmed |
| All results | Every sample was the same value |
Midpoint is halfway along the displayed range. It is not a mean, a median, or a P50 — it is a position on the axis, and treating it as a statistic will mislead you.
Event histograms — a probability that resolves 0 or 1 — leave their two outcomes unnamed, because there is nothing to label.
The numbers are measured over everything
The trimming is presentation only.
Min, max, mean and standard deviation are measured over every sample, including the ones outside the chart. So are the percentiles. Nothing clips the samples themselves, and nothing clips a stored run.
If the mean sits outside the visible axis, that is not an inconsistency — that is what a heavily skewed distribution looks like.
When a node states its value exactly
A node whose value the model already knows reports that value rather than an estimate of it. A binary probability, a probability base rate, and a binary Metaculus question all state their effective figure — after predictions and overrides resolve.
Why this matters
Sampling a node you already know the answer for estimates a number you have. At ten thousand draws, that estimate misses often enough to show an authored 50% as 49% — which looks like a bug and is really just sampling error. Stating the known value removes the confusion.
The samples themselves are untouched. Every formula reading that node still draws from them and keeps its own honest sampling error, and the spread, extremes and histogram are still measured against those draws.
A Uniform distribution likewise reports the exact mean its parameters define, rather than a sampled approximation of it — see Distributions.
A formula's mean is never substituted. What a formula computes is the thing the simulation exists to find out.
Reading a blocked node
A blocked node shows no result at all — not a stale value from before, and not a deterministic placeholder. Its reason is at the top of its inspector and in the diagnostics panel, not on the card.
That is deliberate: a number on a card should always be a number the model currently stands behind.