# HepG2 perturbation-delta measurement reliability audit

- AnnData: `results/state_readiness/GSE264667_hepg2_hvg2000_state_ready.h5ad`
- Eligible targets: 2,160
- HVG response features: 2,000
- Cell split-half repeats: 5

| replicate_definition | n_rows | n_targets | median_rmse | median_cosine | q25_cosine | q75_cosine | median_sign_concordance | statistical_unit |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| cell_split_half | 10800 | 2160 | 0.135929 | 0.331424 | 0.120996 | 0.693082 | 0.569 | target-level replicate delta |
| batch_split | 2160 | 2160 | 0.137286 | 0.33331 | 0.11071 | 0.689825 | 0.5685 | target-level replicate delta |
| guide_pair | 133 | 133 | 0.14349 | 0.00604523 | -0.0611808 | 0.107503 | 0.5025 | target-level replicate delta |

Cell split-half is the broadest technical reliability view. Batch-split preserves batch separation but is sensitive to sparse cells within a target-batch. `sgID_AB` encodes guide-pair/construct identity rather than separable single-guide effects; guide-pair reproducibility is therefore available for only a subset and is reported as exploratory.

The target, not the cell, is the inferential unit. Cells are used only to estimate each replicate delta. The reliability weight is descriptive and was not used to retrain or tune the released model.