What problem does it solve? Relabeling a single-cell dataset is a campaign spanning many sessions, rebuilds, and re-embeddings, and ad-hoc annotation loses selections, rationale, and reproducibility between rounds. This Skill scaffolds one persistent marimo notebook that accumulates lasso, rule, and marker selections across rounds, records every call in a manifest, and deterministically resolves saved selections into a cleaned, relabeled roster. ## Core Features & Use Cases - Multi-round selection instrument: One marimo + jscatter notebook serves an entire campaign, with lasso, rule-based, and marker selections over per-lineage re-embeddings, and a tri-state selection model that persists origin and scope. - Selection manifest and history: Every save writes a CSV plus a manifest row capturing action, new label, priority, rationale, and status, with append-by-default semantics and _history/ snapshots so earlier work stays recoverable. - In-kernel evidence: Mann-Whitney DE tests, runtime gene streaming from CSC sidecars, marker co-detection scatter plots, and genomic track pileups for paired-ATAC datasets, all computed while the lasso is still live. - Deterministic downstream resolution: A resolution stage turns the manifest into per-barcode verdicts, and a cleaning stage applies debris, doublet, and relabel rules to produce a roster with coarse and fine labels. - Use Case: A researcher cleaning a snMultiome atlas runs round 3 of a T-cell lineage relabel: they lasso a suspicious cluster, run an in-kernel DE check against the rest of the lens, save the selection as 'drop' with rationale, and the cleaning stage later emits the updated roster. ## Quick Start Ask the assistant to scaffold a notebook-annotation campaign for your dataset, first telling it what is being relabeled, against which label space, and which round and starting roster this campaign begins from.
