AnnData slot semantics

cp-anndata-validator assumes the following conventions when validating an .h5ad file. It never writes to any of these slots.

Slot	Expected content	Checked by
.X	The primary feature matrix: one row per observation (cell/well/treatment), one column per feature. Must be numeric; sparse or dense; NaN/Inf are flagged but not rejected outright; its shape must equal (n_obs, n_vars).	checks/matrix.py (MATRIX001, MATRIX002, MATRIX003), checks/structure.py (STRUCT002)
.obs	One row per observation, with identifier columns (plate/well/site/cell/perturbation/perturbation modality) and annotation columns (control type, batch, source). Matched via schema aliases, not fixed names.	checks/identifiers.py, checks/annotations.py, checks/metadata.py
.var	One row per feature; var_names are expected to be prefixed by a declared compartment (Cells_, Cytoplasm_, Nuclei_, Image_, ...) followed by a recognized measurement family (AreaShape_, Intensity_, Texture_, ...).	checks/features.py, checks/structure.py (INDEX002)
.uns	Free-form, dataset-level metadata: schema_id, schema_version, processing_stage, layer_processing_stages, licence/license, experiment, image_provenance, segmentation_provenance, feature_extraction_provenance, aggregation (for well/treatment profiles).	checks/schema_meta.py, checks/metadata.py, checks/provenance.py, checks/aggregation.py, checks/matrix.py (SLOT001, SLOT003)
.obsm	Per-observation arrays (embeddings, spatial coordinates, ...). First dimension must equal n_obs.	checks/matrix.py (SLOT002)
.varm	Per-feature arrays (loadings, ...). First dimension must equal n_vars.	checks/matrix.py (SLOT002)
.layers	Alternative versions of .X (raw counts, normalized values, ...). Every layer's shape must equal .X's shape, and each layer should have its own declared processing stage (see below) since uns['processing_stage'] alone only describes .X.	checks/matrix.py (MATRIX004, SLOT003)

Suggested .uns metadata block shapes

None of these are enforced beyond presence/basic shape (the validator does not validate arbitrary nested schemas for .uns sub-blocks), but checks look for these specific shapes:

adata.uns["schema_id"] = "generic-cell-painting"
adata.uns["schema_version"] = "0.2.1"
adata.uns["processing_stage"] = "normalized"  # raw | normalized | aggregated | ...
adata.uns["licence"] = "CC0-1.0"

adata.uns["experiment"] = {"instrument": "...", "protocol": "..."}
adata.uns["image_provenance"] = {"microscope": "...", "illumination_correction": True}
adata.uns["segmentation_provenance"] = {"tool": "CellProfiler", "version": "4.2.6"}
adata.uns["feature_extraction_provenance"] = {"tool": "CellProfiler", "version": "4.2.6"}

# only if .layers holds data at a different processing stage than .X:
adata.uns["layer_processing_stages"] = {"counts": "raw"}

# well-level / treatment-level profiles only. `source_level` is the profile
# level rows were aggregated *from* (what makes plate/well optional at
# treatment level -- see IDENT006 in checks.md):
adata.uns["aggregation"] = {
    "method": "median",
    "replicate_count": 4,
    "source_level": "well",
}

See Rule catalogue for exactly which rule code each block's absence triggers, and Schemas for how .obs/.var columns are resolved via aliases (a separate mechanism from these .uns blocks).