Logbook · Entry 03Ensemble interpretation2 min read

Column ceilings can hide ensemble spread

One strong REPS hour showed all 21 members in near-perfect agreement: boundary-layer tops within 16 m of each other, sitting at the very top of the published column. Agreement that tight among 21 model realizations would be a strong signal — if it came from the atmosphere. It came from the column.

What the 21 members actually shared

Each member passes through the complete nonlinear derivation as its own atmosphere before percentile aggregation. Averaging source temperature, moisture, and heat flux first would manufacture a column no member predicted. (The current percentile shape, contributor counts, full-dropout form, and circular wind rule live in Ensemble values.) So 21 members agreeing to within 16 m meant 21 separate derivations landing in the same place.

The place they landed was the input’s edge. The published column stopped at 700 hPa, about 3,150 m, and the hour was strong enough that every member’s derivation ran into that ceiling. When every atmosphere is truncated at the same height, every boundary-layer top comes out at that height. The narrow spread described the shared input ceiling, not the weather.

Raise the ceiling, and the disagreement reappears

Adding 500 hPa as computational headroom raised the ceiling to about 5,700 m. For the same hour, the boundary-layer spread widened from 16 m to 205 m and the usable-lift spread widened to 425 m — the members had disagreed all along, and the column had been clipping the disagreement off. Hours that had already terminated below 700 hPa remained bit-for-bit unchanged, the control confirming that the headroom changed nothing except where truncation bites. And because a later atmosphere could reach even the higher ceiling, ceiledMembers preserves the warning rather than letting false agreement masquerade as confidence again.

See the two signals side by side below — genuine spread, and the censored profile that declares its own truncation.

Spread and column censoring are different signals

A narrow percentile band can mean member agreement, or it can mean every member reached the same source-column ceiling.

Three ensemble Meteograms: tightly grouped members, widely spread members, and a profile whose boundary-layer percentiles are all capped at the source-column top.

A tightly grouped synthetic ensemble

Closely grouped member inputs produce narrow percentile bands while preserving circular wind direction across north.

900m 2953ft 1599m 5245ft 2298m 7538ft 2996m 9831ft 3695m 12123ft 4394m 14416ft 12 13 14 15 16 17

A widely spread synthetic ensemble

Broader source perturbations produce visibly wider percentile bands without changing member accounting or derivation authority.

900m 2953ft 1599m 5245ft 2298m 7538ft 2996m 9831ft 3695m 12123ft 4394m 14416ft 12 13 14 15 16 17

A synthetic ensemble capped by its column

A deliberately shallow published column marks derived height members that reach the highest available level instead of presenting the ceiling as a resolved top.

900m 2953ft 1599m 5245ft 2298m 7538ft 2996m 9831ft 3695m 12123ft 4394m 14416ft 12 13 14 15 16 17
Every band is serialized from the scenario's published percentile blocks. The censored profile declares ceiledMembers rather than asking line width to carry that warning alone.Units altitude m and ft · percentile bands p25–p75 · time UTC

Spread is not probability

Member fraction measures agreement among model realizations. “18 of 21” does not establish an 86% chance of usable lift. Probability requires calibration against archived forecasts and observations; this project does not yet have that evidence.

The forecast model feed reference records the current source levels and provider paths. ECCC documents the underlying system on its REPS open-data page.