Consensus sweep job
The kNN graph is built once per (graph type × n_neighbors) and every resolution × seed reruns Leiden on it. Grid points are nearly free; each graph type or neighbourhood size costs a graph build.
Grain window job
Narrows which runs are pooled, by realised cluster count. Changes what every threshold below means, so read the coverage verdict before trusting it.
Ensemble audit free
ARI of each graph type against the pooled cut, scored at three probe thresholds. A slice at half the others' agreement is evidence about that weighting, not noise to average in.
Embedding job
Needed for the Map tab. Uses this level's space.
Threshold scan free
Each point is one fcluster on the cached linkage. Shaded bands are plateaus — runs of thresholds giving the same cluster count, i.e. gaps in the dendrogram. Cut in the middle of one, not at its edge. Click the chart to set the threshold.
Merge support job
Per resolution band, the fraction of runs whose modal labels kept the two groups together. all-band = a type; fine-only = a subtype split that exists only because the fine runs could make it; unsupported = nothing in the ensemble backs this merge, so your threshold should not sit above it. Supports are sampled estimates — 0.48 against 0.52 is a tie.
Flow across the ladder free
Cells flowing between adjacent cuts, coarse on the left. Flows are matched by membership, never by cluster id — ids are renumbered to 0..k-1 after size filtering and are not stable across levels. The grey band is unassigned: coarsening only grows clusters, so it drains in one direction and never refills.
Embedding free
Coloured by stability, low scorers should concentrate along interdigitated boundaries and continuous streaks. Scattered uniformly instead, the instability is not about boundaries and the cut is not the thing to adjust.
Depth and metadata free
One column's distribution per cluster on a shared binning — the view that says whether a split is laminar. Raw counts show which clusters dominate; normalized shows shape, so a small cluster with a sharp laminar position is invisible in the first and obvious in the second.
Cluster × feature free
Z-scored feature means per cluster, ordered by a one-way F statistic across clusters — which features separate these clusters, not which vary most. Cells are standard deviations from the cohort mean.
Feature against column free
The follow-through on the heatmap: a cluster mean says the feature differs, this says whether it differs as a shift, a spread, or a gradient — three different claims about what the cluster is.
Gap or cut job
A tool that only offers "cut into k clusters" will always answer that the data is discrete. This asks the prior question, per cluster pair, on four independent legs: a dip test on the projected separation, graph connectivity, a density valley, and the consensus co-clustering profile. The verdict is a majority of the legs that expressed an opinion — a summary, not a result. Re-check it at another n_neighbors before acting on it.
Parametrize a continuum job
The alternative to a label, not a decoration on one. Fits a principal curve through the selected clusters and returns a coordinate in [0,1]. Two warnings are part of the contract and are shown, not logged: the dimension gate (a curve through genuinely higher-dimensional structure compresses it into an artifact) and the nuisance tripwire (truncation manufactures gradients — a coordinate that tracks reconstruction quality is not biology acquiring a name).
Lock in this level
Appends to the ledger, attaches the labels, and carves the child masks — which open as new levels. The threshold is recorded together with the grain window and size floor it was chosen under, because it does not mean anything without them.