sei_pertrack.npz: Repair build_config storage shape to (1,) on the five layers_per_row-stamped artefacts. The first stamp wrote a 0-d array, which raises IndexError in readers that do build_config[0]. CDF matrices untouched throughout.
legnet_pertrack.npz: Repair build_config storage shape to (1,) on the five layers_per_row-stamped artefacts. The first stamp wrote a 0-d array, which raises IndexError in readers that do build_config[0]. CDF matrices untouched throughout.
epinformerseq_pertrack.npz: Repair build_config storage shape to (1,) on the five layers_per_row-stamped artefacts. The first stamp wrote a 0-d array, which raises IndexError in readers that do build_config[0]. CDF matrices untouched throughout.
chrombpnet_pertrack.npz: Repair build_config storage shape to (1,) on the five layers_per_row-stamped artefacts. The first stamp wrote a 0-d array, which raises IndexError in readers that do build_config[0]. CDF matrices untouched throughout.
cherimoya_pertrack.npz: Repair build_config storage shape to (1,) on the five layers_per_row-stamped artefacts. The first stamp wrote a 0-d array, which raises IndexError in readers that do build_config[0]. CDF matrices untouched throughout.
sei_pertrack.npz: Stamp layers_per_row on the five oracles that shipped without it (v0.7.1). Derived by asking classify_track_layer -- the same function the query path calls -- so the stored array cannot disagree with what downstream code computes. ChromBPNet is the case that mattered: 753 rows over 2 layers (chromatin_accessibility 9, tf_binding 744), where code keying on the array previously had to re-derive the layer from the track-id string. CDF matrices are byte-identical: append-in-place, no rebuild.
legnet_pertrack.npz: Stamp layers_per_row on the five oracles that shipped without it (v0.7.1). Derived by asking classify_track_layer -- the same function the query path calls -- so the stored array cannot disagree with what downstream code computes. ChromBPNet is the case that mattered: 753 rows over 2 layers (chromatin_accessibility 9, tf_binding 744), where code keying on the array previously had to re-derive the layer from the track-id string. CDF matrices are byte-identical: append-in-place, no rebuild.
epinformerseq_pertrack.npz: Stamp layers_per_row on the five oracles that shipped without it (v0.7.1). Derived by asking classify_track_layer -- the same function the query path calls -- so the stored array cannot disagree with what downstream code computes. ChromBPNet is the case that mattered: 753 rows over 2 layers (chromatin_accessibility 9, tf_binding 744), where code keying on the array previously had to re-derive the layer from the track-id string. CDF matrices are byte-identical: append-in-place, no rebuild.
chrombpnet_pertrack.npz: Stamp layers_per_row on the five oracles that shipped without it (v0.7.1). Derived by asking classify_track_layer -- the same function the query path calls -- so the stored array cannot disagree with what downstream code computes. ChromBPNet is the case that mattered: 753 rows over 2 layers (chromatin_accessibility 9, tf_binding 744), where code keying on the array previously had to re-derive the layer from the track-id string. CDF matrices are byte-identical: append-in-place, no rebuild.
cherimoya_pertrack.npz: Stamp layers_per_row on the five oracles that shipped without it (v0.7.1). Derived by asking classify_track_layer -- the same function the query path calls -- so the stored array cannot disagree with what downstream code computes. ChromBPNet is the case that mattered: 753 rows over 2 layers (chromatin_accessibility 9, tf_binding 744), where code keying on the array previously had to re-derive the layer from the track-id string. CDF matrices are byte-identical: append-in-place, no rebuild.
borzoi_pertrack.npz: Unified background rebuild, schema 4 (build_campaign 2026-08-06). Fixes the reservoir-thinning defect: a uniform m-of-N subsample retained the population maximum with probability exactly m/N, so 9 of 19 (oracle, layer) reservoir pairs had understated ceilings -- AlphaGenome gene_expression by up to 8.3x. effect and summary layers are now exact; perbin keeps an exact top-K tail. Cherimoya is the 5-fold CATv1 ensemble. Activity provenance is derived per oracle rather than stamped uniformly.
enformer_pertrack.npz: Unified background rebuild, schema 4 (build_campaign 2026-08-06). Fixes the reservoir-thinning defect: a uniform m-of-N subsample retained the population maximum with probability exactly m/N, so 9 of 19 (oracle, layer) reservoir pairs had understated ceilings -- AlphaGenome gene_expression by up to 8.3x. effect and summary layers are now exact; perbin keeps an exact top-K tail. Cherimoya is the 5-fold CATv1 ensemble. Activity provenance is derived per oracle rather than stamped uniformly.
alphagenome_pertrack.npz: Unified background rebuild, schema 4 (build_campaign 2026-08-06). Fixes the reservoir-thinning defect: a uniform m-of-N subsample retained the population maximum with probability exactly m/N, so 9 of 19 (oracle, layer) reservoir pairs had understated ceilings -- AlphaGenome gene_expression by up to 8.3x. effect and summary layers are now exact; perbin keeps an exact top-K tail. Cherimoya is the 5-fold CATv1 ensemble. Activity provenance is derived per oracle rather than stamped uniformly.
cherimoya_pertrack.npz: Unified background rebuild, schema 4 (build_campaign 2026-08-06). Fixes the reservoir-thinning defect: a uniform m-of-N subsample retained the population maximum with probability exactly m/N, so 9 of 19 (oracle, layer) reservoir pairs had understated ceilings -- AlphaGenome gene_expression by up to 8.3x. effect and summary layers are now exact; perbin keeps an exact top-K tail. Cherimoya is the 5-fold CATv1 ensemble. Activity provenance is derived per oracle rather than stamped uniformly.
chrombpnet_pertrack.npz: Unified background rebuild, schema 4 (build_campaign 2026-08-06). Fixes the reservoir-thinning defect: a uniform m-of-N subsample retained the population maximum with probability exactly m/N, so 9 of 19 (oracle, layer) reservoir pairs had understated ceilings -- AlphaGenome gene_expression by up to 8.3x. effect and summary layers are now exact; perbin keeps an exact top-K tail. Cherimoya is the 5-fold CATv1 ensemble. Activity provenance is derived per oracle rather than stamped uniformly.
sei_pertrack.npz: Unified background rebuild, schema 4 (build_campaign 2026-08-06). Fixes the reservoir-thinning defect: a uniform m-of-N subsample retained the population maximum with probability exactly m/N, so 9 of 19 (oracle, layer) reservoir pairs had understated ceilings -- AlphaGenome gene_expression by up to 8.3x. effect and summary layers are now exact; perbin keeps an exact top-K tail. Cherimoya is the 5-fold CATv1 ensemble. Activity provenance is derived per oracle rather than stamped uniformly.
epinformerseq_pertrack.npz: Unified background rebuild, schema 4 (build_campaign 2026-08-06). Fixes the reservoir-thinning defect: a uniform m-of-N subsample retained the population maximum with probability exactly m/N, so 9 of 19 (oracle, layer) reservoir pairs had understated ceilings -- AlphaGenome gene_expression by up to 8.3x. effect and summary layers are now exact; perbin keeps an exact top-K tail. Cherimoya is the 5-fold CATv1 ensemble. Activity provenance is derived per oracle rather than stamped uniformly.
legnet_pertrack.npz: Unified background rebuild, schema 4 (build_campaign 2026-08-06). Fixes the reservoir-thinning defect: a uniform m-of-N subsample retained the population maximum with probability exactly m/N, so 9 of 19 (oracle, layer) reservoir pairs had understated ceilings -- AlphaGenome gene_expression by up to 8.3x. effect and summary layers are now exact; perbin keeps an exact top-K tail. Cherimoya is the 5-fold CATv1 ensemble. Activity provenance is derived per oracle rather than stamped uniformly.
