Imported from francescorubbo/jumpbench (
AGENTS.md). Install upstream withnpx skills add francescorubbo/jumpbench. Copyright stays with the author.
Agent notes
This repo generates controllable timm embeddings and scores CRISPR PA
after simple_pca100. JUMP_lite remains the
source of truth for their numbers. jumpbench compare --mode paper_as_published
is tagged unfair on purpose.
Ignore the paper CRISPR PA NAP 0.815. It is JUMP_lite’s PA×PC-selected
full-CRISPR config, not this study’s endpoint. We rank CRISPR-subset PA
(mean NAP) against our own arms. Run1 is a smaller slice, so 0.815 is even
less comparable. Do not write paper_nap / delta_vs_paper into result JSON
or treat 0.815 as a target.
Ranking fairness
Living preregistration: docs/hypotheses.md.
When an analysis supports or kills a claim, update that hypothesis’s Status
and Decision in place (date + evidence). Add a new hypothesis instead of
silently expanding an old one. Do not mark an observational hypothesis
falsified because we chose not to run the experiment.
Study constraints
Do not propose work that violates these:
- Cellpose masks exist only for the JUMP-lite 4-site subset.
--crop cell_fixed/cell_bboxis 4-site-only. All-FOV (--sites all) runs are grid tiles only. - Endpoint is CRISPR PA after
simple_pca100(PCA-100, plate negcon z-score; or a declared DL sweep). No PC, no MOTIVE, no 11-task Figure 5 mean.paper_dl_defaultis a comparator, not the readout. - New representations come only from
jumpbench embed --model timm. Do not re-extract CellProfiler,cp_measure, or MorphEM from pixels. Frozen comparators: paper headline numbers and assembled CPG CellProfiler profiles already on disk. Swappingmodels.timm.architecture(EfficientNet vs a DINOv2-class ViT) is in-scope for H12; that is not a re-run of paper DINOv2. H13 has a declared Raw-vs-MQ timm analogue (Wave R: cell-crop EfficientNetV2-XL on streamed Orig TIFF vs streamed JUMP-litejpegxl_lossy_mq.zarr). That is not a Table 1 four-codec restudy; do not start HQ/D20 or re-embed paper DL families.
Campaign runbook: docs/campaign.md.
Pointers
- docs/protocol.md — PA/PC protocol and fairness modes
- configs/models.yaml — channel recipes, tile sizes, preprocess