Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261001T200638Z-extra-realism-metrics/diffusion/dataset/server_1/model8/plans/paired_finetune_intradataset_20260927/scores_extra/README.md

@@ -0,0 +1,205 @@
+# scores_extra: reviewer-expected realism and diversity metrics (added 2026-10-01/02)
+
+Written by `tools/extra_metrics.py` (CPU only; tests: `tests/test_extra_metrics.py`; launcher: `tools/run_extra_unit.sh`;
+tables: `tools/collect_extra_metrics.py`, also run by `extra_metrics.py --collect`). Code versions per unit: `_code/`.
+Nothing under `scores/` or `generated/` was modified. Every unit re-reads one existing `score_ft.py` scoring call
+(`scores/<ds>/fewshot_n5/draw<j>` + `draw<j>_scsemi5`, or `scores/<ds>/pooled`) and adds four metric families next to
+its precision/coverage. Tables: `TABLES.md`.
+
+## Scope
+
+| design | units | arms (arm names in `summary.json`) |
+|---|---|---|
+| (a) 5-donor few-pair | 5 cohorts × 5 draws | `zs_raw` (zero-shot, real bulk), `zs_cal` (zero-shot + calibration, α from the same 5 pairs), `scsemi5` (scSemiProfiler, 5 donors), `lora5` (LoRA, 5 donors); reference `zs_pseudo` (zero-shot, sc pseudobulk input) |
+| (b) 5-fold pooled | 5 cohorts × 1 pooled call | `zs__raw`, `zs__cal`, `pseudo_full__cal` (sc-only fine-tune + calibration), `real_lora__raw` (LoRA, all training pairs); reference `zs__pseudo` |
+
+Controls in every unit: `real_R2` (the scorer's own real-vs-real control: R1 vs R2, same donors, same n per donor)
+and `real_R3` (out-of-sample control, see below).
+
+## Inputs re-used exactly (checked per unit, `repro.json`)
+
+- Reference: `scoring_cache/<ds>/reference.npz` (sha = the scored call's `reference_sha256`), `score_ft.make_view` on
+  the call's scored donors → PCA-50 (randomized, random_state 20260925) fitted on the scored donors' **R1** cells,
+  k = 10 radii. PCA explained-variance sum and the R2 control must equal the call's `summary.json`.
+- Generated cells: the scored NPZ (sha256 = `per_arm.npz_sha256`, generation receipt `output_sha256` verified via
+  `score_ft.open_arm`; NPZs moved to server_2 were rsynced to a temporary scratch copy, sha-checked against the
+  receipt, the `.moved` marker and `summary.json`, and deleted after the unit/cohort). One Poisson read-out of the
+  whole NPZ (seed 20260916, `poisson_integerise`, chunk 4,096), scored donors' rows, scoring genes.
+- P/C subsample: `ft_common.generated_subsample` (n = the donor's R1 size, `default_rng([20260926, key])`); the rows
+  must equal `arms/<arm>/generated_cells.parquet` (`npz_row`, `in_pc_subsample`).
+- Depth matching: `stage_b_embed.depth_match` run pairwise (real set vs query set, fresh `rng(20260929)`), then
+  `stage_b_embed.project` with the un-thinned PCA; radii recomputed on the thinned real set (as `score_ft.depth_matched`).
+- **Reproduction check:** precision and coverage at k = 10, pooled and per donor, raw and depth matched, for every
+  arm and the R2 control, recomputed and compared with `summary.json` / `per_donor.csv` (max |difference| in
+  `repro.json` and `*_reproduction_check.csv`).
+
+## Real side: R1 (scorer space) and R2 (out-of-sample)
+
+The scorer's PCA-50 is fitted on R1, so R1 points are in-sample and generated / R2 points are projected
+out-of-sample. Precision/coverage tolerate this (R1 vs R2 coverage ≈ 1), but a classifier separates in-sample from
+projected points: in Giroux (1,650 R1 cells, 5,563 genes) the random forest separates R1 from R2 with AUC 0.93.
+Families 1 and 2 are therefore reported twice, with everything else identical:
+
+- `real_side = R1`: real = R1 (the scorer's reference cells, its fitted embedding). Control `real_R2` = R1 vs R2.
+- `real_side = R2`: real = R2 (the scorer's held-out cells, projected like the generated cells). Control `real_R3` =
+  R2′ vs R3, where R3 = `perm[2n:3n]` of the scorer's own donor-keyed permutation (`F.reference_split`: R1 =
+  `perm[:n]`, R2 = `perm[n:2n]`, both rebuilt from the observed cells and checked equal to `reference.npz`), and
+  R2′ = `perm[n:n+m]` with m = min(n, |R3|) so both control sides have the same size per sample (samples with fewer
+  than 3n cells are listed in `repro.json` → `oos_r3.short_samples`). Depth matching is run on the (R2, query) pair.
+  **This is the space to read the classifier tests and MMD² in**; R1-side values are kept because they use exactly
+  the scorer's reference set.
+
+Precision/coverage (family 4) and E_N (family 3) are reported once (R1 side for P/C, its own space for E_N).
+
+## Definitions and parameters
+
+Embedding for families 1, 2, 4: per-cell CP10K over the dataset's scoring genes, log1p, the scorer's PCA-50 (above).
+"raw" = no depth matching (the deck's headline P/C); "dm" = depth matched. Seed of every new random stream: 20261001.
+
+1. **Classifier two-sample tests** (Lopez-Paz & Oquab 2017; scGAN, Marouf et al. 2020; scDiffusion, Luo et al. 2024).
+   Pool real and generated embeddings (equal n per donor: the donor's R1 size, ≤ 150), label real = 1. Stratified
+   5-fold CV (shuffle, seed 20261001); per fold fit, predict the held-out fold's class-1 probability, ROC AUC; report
+   the mean (and SD) over folds.
+   - `c2st_knn_auc`: `KNeighborsClassifier(n_neighbors=5)` (sklearn default, as scDiffusion's kNN test).
+   - `c2st_rf_auc`: `RandomForestClassifier(n_estimators=300, max_features="sqrt", max_depth=None)` (scGAN's fully grown
+     Gini trees with √p features; 300 instead of scGAN's 1,000 trees for run time).
+   - `nn1_acc`: leave-one-out 1-NN accuracy (scLDM, Palla et al. 2025, eq. 32): each pooled point's label is
+     predicted by its nearest other point (Euclidean, brute force float64); also `nn1_acc_real` / `nn1_acc_gen`
+     (accuracy on the real resp. generated points). 0.5 = indistinguishable, 1 = separable.
+   AUC and 1-NN accuracy: closer to 0.5 is better.
+2. **MMD²** (Gretton et al. 2012, unbiased estimator eq. 3): Gaussian RBF k(x, y) = exp(−‖x − y‖² / (2σ²)).
+   Bandwidth: median heuristic m = median pairwise Euclidean distance among ≤ 4,000 random R1 cells of the unit
+   (same m for every arm, mode and real side of the unit; `mmd_bandwidth_m`). `mmd2_bw0.5`, `mmd2_bw1`, `mmd2_bw2`:
+   σ = m/2, m, 2m (scGAN's multiples); `mmd2_sum`: the sum of the three kernels (scGAN's multi-scale kernel).
+   Unbiased MMD² can be slightly negative for identical distributions.
+3. **E_N, scSemiProfiler's normalized error** (Wang et al., *Nat Commun* 15:5989, 2024, Methods eq. 17–18;
+   `scsemi/vendor/scSemiProfiler/scSemiProfiler/utils.py::get_error`, lines 1740–1935, and `errorcurve`).
+   - Space (theirs): per arm, log1p CP10K of real and generated cells, **PCA-100 fitted on real and generated
+     together** (exact PCA via eigh of the covariance; equals sklearn PCA up to component signs, tested).
+   - Dif_sc(a, b) (eq. 17, K = 1) = (Σ_{cells of a} d(cell, NN in b) + Σ_{cells of b} d(cell, NN in a)) / (N_a + N_b).
+     `distance = sq`: squared L2, as their code computes it (faiss `IndexFlatL2` returns squared distances) — the
+     headline; `distance = l2`: Euclidean, as the paper's text writes it.
+   - LB: random half split of the whole PCA matrix (real + generated cells of all scored samples), symmetric 1-NN
+     between the halves. UB: for each sample a uniformly random partner sample **j ≠ i** (their code draws
+     `randint(0, P)`, which can pair a sample with itself), Dif_sc of the two samples' real cells, averaged.
+   - Per sample: Dif_i = Dif_sc(real_i, generated_i), `en_i` = (Dif_i − LB) / (UB − LB). Unit:
+     `en` = (mean_i Dif_i − LB) / (UB − LB) (eq. 18 / `errorcurve`).
+   - **Representatives:** scSemiProfiler sets a representative's Dif to LB (eq. 18 second line). Our scored samples
+     never contain a representative: the 5 training donors of a few-pair draw are not scored in any arm, and in the
+     pooled design every sample is a cross-fitted test sample. No sample is set to LB for any arm. For `scsemi5` the
+     column `en_reps_at_lb` gives the value scSemiProfiler's convention would report with the same LB/UB:
+     `en × n_test / (n_test + n_representative_rows)` (representative rows = all usable rows of the 5 training donors,
+     `data/<ds>/fewshot_n5/DRAWS.json → n_train_rows`).
+   - Cell sets: `cellset = all` (their setting): every observed QC cell of each scored sample vs all its generated
+     cells (1,500); LB estimated from 20,000 random query cells per direction against the complete other half
+     (unbiased for their mean; `lb_exact = False`); control `real_R2` here = a random within-sample half split of the
+     observed cells (half A as real, half B as generated; `F.donor_rng(20261001, sample)`). `cellset = pc`: the P/C
+     cells (R1 vs the generated subsample, equal n per sample), exact LB; control `real_R2` = R1 vs R2.
+   - Equivalence test (`test_en_matches_get_error`): on a toy project, scSemiProfiler's own `get_error` (run in the
+     frozen scSemiProfiler `runtime_env`) and our implementation agree to rtol 2e-5 (UB, LB, semi and naive error, E_N).
+4. **k sensitivity of precision / coverage** (Kynkäänniemi et al. 2019; Naeem et al. 2020): the scorer's
+   `stage_b_embed.knn_radius` / `precision_coverage` with k = 3, 5, 10 (k = 10 = the reported values, reproduced).
+   `density_k*` is Naeem's density rescaled to the k used. `expected_coverage_identical_k*`: Naeem eq. 9,
+   1 − Π_{i=1..k}(N − i)/(M + N − i), the expected coverage of a perfect generator at these sample sizes.
+   Note: Kynkäänniemi precision has no closed form and falls with k even for identical distributions (50-D
+   Gaussians, n = 2,000: 0.62 / 0.74 / 0.86 at k = 3 / 5 / 10), so compare arms with the real-vs-real control row.
+
+## Results in brief (2026-10-02; full tables `TABLES.md`, numbers `*_headline.csv`, `*_wide.csv`)
+
+- **Reproduction:** k = 10 precision/coverage, pooled and per donor, raw and depth matched, equal to `score_ft.py`
+  in all 30 units (max |diff| 1.1e-16); scored rows and P/C subsample identical in every arm.
+- **Controls:** out-of-sample real-vs-real (R2 vs R3): AUC kNN / RF 0.49–0.51, 1-NN 0.50–0.51, MMD² ≈ 0 in every
+  cohort. Scorer-space R1 vs R2: RF AUC 0.56–0.59 (ImMicro, sepsis, Imperial), 0.77–0.81 (Ji), 0.84–0.94 (Giroux) —
+  the in-sample PCA artefact that motivates the R2 headline space.
+- **LoRA vs comparators** (headline spaces, no depth matching): pooled 5-fold — better than zero-shot raw, zero-shot
+  + calibration and sc-only fine-tune (+ calibration) on every new metric in all 5 cohorts. Few-pair — better than
+  scSemiProfiler and zero-shot raw in every draw on every metric except E_N on the P/C cells (Giroux 0/5, Ji 4/5 vs
+  scSemiProfiler); vs zero-shot + calibration (draws where LoRA is better), exceptions in Imperial (E_N all cells
+  2/5, E_N P/C cells 2/5, precision k3/5/10 3/5, MMD² 4/5), sepsis (precision k5 4/5, k10 3/5) and Giroux (MMD² 4/5,
+  precision k5/k10 4/5). AUCs (kNN, RF),
+  1-NN accuracy and coverage at every k: LoRA better in 75/75 few-pair comparisons and 15/15 pooled comparisons.
+- **E_N, scSemiProfiler vs LoRA (few-pair, all cells, their squared L2):** LoRA lower in 25/25 draws (L2: 24/25;
+  un-normalized Dif: 25/25), per sample in 96.5 % of 1,790 sample-draws. Means: ImMicro 3.65 vs 1.55, sepsis 3.83 vs
+  1.46, Imperial 2.04 vs 1.28, Giroux 3.80 vs 1.99, Ji 4.27 vs 2.60. With scSemiProfiler's own convention
+  (representatives at LB) its value would be 3.48 / 3.60 / 1.86 / 1.71 / 3.26 — below LoRA's 1.99 in Giroux, where 12–14
+  of 24 rows are representatives (LoRA is never credited that way).
+- **E_N caveats** (properties of eq. 18, not of our code; the toy test matches `get_error`):
+  1. Every generated arm has E_N > 1 (real-vs-real half split 0.07–0.38): at the single-cell 1-NN level, generated
+     cells are farther from their own sample's real cells than another real sample is.
+  2. LB and UB are per arm. LB pools the arm's own generated cells, so it can approach or exceed UB for off-manifold
+     arms (Ji zero-shot raw few-pair E_N 33 ± 16). UB ≤ LB (E_N undefined, negative) occurred only for the P/C-cell
+     version in Ji: zero-shot raw in all 5 draws and the pooled call, zero-shot + calibration in draw 1.
+     Undefined units (UB ≤ LB) are flagged † in `TABLES.md` and rank last.
+  3. On the small P/C cell set, scSemiProfiler's LB is low because its cells for targets that share a representative
+     are near-copies: median cross-sample 1-NN distance of generated cells (scorer PCA-50, draw 0) ImMicro 4.6 vs LoRA
+     6.9, sepsis 4.7 vs 7.9, Giroux 9.5 vs 11.5. This flips P/C-cell E_N in Giroux although its Dif is higher (684 vs
+     618). The all-cells version (their setting) is dominated by real cells and is the one to report.
+- **k sensitivity:** coverage and precision fall with smaller k for every set including the real control (e.g. ImMicro
+  control C 1.000 / 0.988 / 0.932, P 0.965 / 0.907 / 0.835 at k = 10 / 5 / 3). The arm ranking by coverage is
+  unchanged at k = 3 and 5 in every cohort except one swap in ImMicro (both designs) between zero-shot + calibration
+  and the zero-shot pseudobulk reference, which differ by < 0.005 at k = 10; LoRA is first at every k.
+- **Run time** (server_1 CPU; few-pair 6 threads, pooled 8 threads; per unit = all arms + controls, all variants):
+  few-pair Giroux 245 s, Ji 330 s, sepsis 1,331 s, ImMicro 1,446 s, Imperial 2,249 s (incl. fetching ~25 GB from
+  server_2); pooled Ji 387 s, Giroux 461 s, sepsis 1,373 s, ImMicro 1,591 s, Imperial 2,449 s. Total 9.2 h of unit
+  time, 5.0 h wall clock on two streams (2026-10-02 00:01–05:00 KST).
+
+## Files
+
+Per unit `scores_extra/<ds>/{fewshot_n5/draw<j>, pooled}/`:
+- `metrics.csv`: one row per set × real_side × mode (families 1, 2 and, for real_side R1, 4).
+- `per_donor.csv`: per set × donor precision/coverage at k = 3/5/10, raw and depth matched (R1 side).
+- `en.csv`: per set × cellset × distance: LB, UB, mean Dif, E_N, `en_reps_at_lb` (scsemi5), sizes, seconds.
+- `en_per_sample.csv`: per set × cellset × distance × sample: n real / generated cells, UB partner, Dif_i, E_N,i.
+- `repro.json`: k = 10 reproduction (pooled + per-donor max |diff|), PCA, R2 control, row/subsample identity, R3 sizes.
+- `manifest.json`: inputs (summary sha256, reference sha, NPZ path + sha), parameters, code sha256, run time.
+
+Collected (`tools/collect_extra_metrics.py`): `{fewshot,pooled}_{metrics,per_donor,en,en_per_sample}.csv` (all units
+concatenated), `{fewshot,pooled}_wide.csv` (one row per unit × set × real side × mode, with E_N, Dif, LB, UB),
+`{fewshot,pooled}_headline.csv` (cohort × arm headline values; few-pair mean, SD, n over draws),
+`{fewshot,pooled}_reproduction_check.csv`, `run_time.csv`, `TABLES.md`. `_code/`: the `extra_metrics.py` versions
+that ran and the unit → version map.
+
+## Draft Methods paragraph
+
+**Additional realism and diversity metrics.** To complement precision and coverage we computed three further
+families of metrics on exactly the cells, embedding and depth matching used for precision and coverage, plus
+scSemiProfiler's own error metric. For every scored unit (one few-pair draw, or the pooled call over all
+cross-fitted 5-fold pairs) the real cells are each sample's reference cells (n = min(150, ⌊cells/2⌋), donor-keyed
+draw) and the generated cells an equal-sized donor-keyed subsample of the Poisson-read generated cells; both are
+log1p-CP10K transformed on the cohort's scoring genes and projected on the 50-component PCA fitted to the
+reference cells. Because that PCA places the reference cells in-sample, two-sample statistics use a second held-out
+set of real cells of the same samples as the real side (same n per sample), and a third held-out set as the
+real-versus-real control. (i) Classifier two-sample tests (Lopez-Paz & Oquab 2017): a k-nearest-neighbour classifier
+(k = 5) and a random forest (300 fully grown trees, √p features per split) were trained to separate real from
+generated cells, and we report the stratified five-fold cross-validated ROC AUC (as in scGAN and scDiffusion), with
+the leave-one-out 1-nearest-neighbour accuracy (as in scLDM); 0.5 indicates indistinguishable sets. (ii) The unbiased
+squared maximum mean discrepancy (Gretton et al. 2012) with a Gaussian kernel whose bandwidth σ is the median
+pairwise distance among real cells, evaluated at σ/2, σ and 2σ and for their sum (as in scGAN). (iii)
+scSemiProfiler's normalized error E_N (Wang et al. 2024, eq. 17–18), computed exactly as their implementation:
+per method, all observed cells and all generated cells of the scored samples were jointly reduced to 100 principal
+components; for each sample the symmetric nearest-neighbour (squared Euclidean) distance between its real and
+generated cells was normalized by a lower bound (random half split of the whole matrix) and an upper bound (real
+cells of randomly paired different samples). Since no scored sample was a representative, no sample was set to the
+lower bound. (iv) Precision and coverage were recomputed at k = 3 and 5 alongside k = 10, with Naeem et al.'s
+expected coverage of a perfect generator as reference. All metrics were also computed after depth matching.
+Recomputed k = 10 precision and coverage matched the reported values exactly in every unit, and our E_N
+implementation reproduced scSemiProfiler's reference code on a test project.
+
+## Citations
+
+- Kynkäänniemi T. et al. Improved precision and recall metric for assessing generative models. NeurIPS 2019.
+- Naeem M. F. et al. Reliable fidelity and diversity metrics for generative models. ICML 2020.
+- Gretton A. et al. A kernel two-sample test. JMLR 13:723–773, 2012.
+- Lopez-Paz D. & Oquab M. Revisiting classifier two-sample tests. ICLR 2017.
+- Marouf M. et al. Realistic in silico generation and augmentation of single-cell RNA-seq data using GANs (scGAN).
+  Nat Commun 11:166, 2020 (doi:10.1038/s41467-019-14018-z).
+- Luo E. et al. scDiffusion: conditional generation of high-quality single-cell data using diffusion model.
+  Bioinformatics 40(9):btae518, 2024 (arXiv 2401.03968).
+- Palma A. et al. Multi-modal and multi-attribute generation of single cells with CFGen. ICLR 2025.
+- Palla G. et al. Scalable single-cell gene expression generation with latent diffusion models (scLDM).
+  arXiv 2511.02986 (2025).
+- Wang J. et al. Advancing large-scale single-cell studies through semi-profiling with deep generative models and
+  active learning (scSemiProfiler). Nat Commun 15:5989, 2024.
+
+All PDFs are in `/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/` (`evaluation-metrics/`,
+`sc-generation/`, `related-works/`, `bulk-to-sc/`).

paper-releases/pbisc-diffusion/updates/20261001T200638Z-extra-realism-metrics/diffusion/dataset/server_1/model8/plans/paired_finetune_intradataset_20260927/scores_extra/TABLES.md

@@ -0,0 +1,471 @@
+# Extra realism / diversity metrics: tables
+
+Written by `tools/collect_extra_metrics.py` (2026-10-01T20:01:48.507974+00:00). Definitions, parameters and citations: `README.md`.
+
+Headline spaces: AUC (5-fold CV), 1-NN accuracy and MMD² use the **out-of-sample real side R2** (same PCA-50, generated cells, subsample as the scorer; control row = R2 vs R3); precision (P) / coverage (C) are the scorer's own (real side R1; control row = R1 vs R2); E_N = scSemiProfiler's normalized error on all cells, squared L2 (control row = within-sample half split). AUC / 1-NN: 0.5 is ideal. † n = units where E_N is undefined (the arm's UB ≤ LB), left out of the mean. No depth matching in the headline tables.
+
+## (a) 5-donor few-pair design: mean ± SD over the 5 draws
+
+Units: 25. k = 10 reproduction: max |diff| vs score_ft.py = 1.1e-16 over 150 unit × set rows (pooled and per donor, raw and depth matched); scored rows and P/C subsample identical in every arm: True.
+
+| cohort | arm | AUC kNN | AUC RF | 1-NN acc | MMD² (σ=m) | MMD² (σ sum) | E_N (all cells) | P k3 | C k3 | P k5 | C k5 | P k10 | C k10 |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| ImMicro | Real vs real control | 0.502 ± 0.001 | 0.496 ± 0.002 | 0.498 ± 0.001 | -2.6e-05 ± 0.0000 | -6.1e-05 ± 0.0000 | 0.382 ± 0.015 | 0.835 ± 0.001 | 0.932 ± 0.002 | 0.907 ± 0.001 | 0.988 ± 0.001 | 0.965 ± 0.001 | 1.000 ± 0.000 |
+| ImMicro | Zero-shot, real bulk | 0.997 ± 0.000 | 0.998 ± 0.000 | 0.981 ± 0.001 | 0.1026 ± 0.0005 | 0.2487 ± 0.0010 | 7.236 ± 0.345 | 0.399 ± 0.009 | 0.087 ± 0.003 | 0.521 ± 0.003 | 0.122 ± 0.003 | 0.677 ± 0.006 | 0.190 ± 0.004 |
+| ImMicro | Zero-shot + calibration | 0.973 ± 0.002 | 0.986 ± 0.001 | 0.909 ± 0.003 | 0.0430 ± 0.0024 | 0.1199 ± 0.0067 | 2.855 ± 0.085 | 0.706 ± 0.014 | 0.346 ± 0.003 | 0.807 ± 0.008 | 0.425 ± 0.005 | 0.904 ± 0.006 | 0.535 ± 0.008 |
+| ImMicro | scSemiProfiler, 5 donors | 0.999 ± 0.000 | 1.000 ± 0.000 | 0.996 ± 0.000 | 0.1865 ± 0.0092 | 0.4991 ± 0.0241 | 3.649 ± 0.046 | 0.236 ± 0.025 | 0.019 ± 0.001 | 0.333 ± 0.034 | 0.025 ± 0.001 | 0.518 ± 0.036 | 0.036 ± 0.002 |
+| ImMicro | **LoRA, 5 donors** | 0.943 ± 0.003 | 0.960 ± 0.002 | 0.839 ± 0.008 | 0.0231 ± 0.0045 | 0.0677 ± 0.0127 | 1.547 ± 0.076 | 0.797 ± 0.017 | 0.564 ± 0.039 | 0.882 ± 0.012 | 0.671 ± 0.045 | 0.952 ± 0.005 | 0.792 ± 0.047 |
+| ImMicro | Zero-shot, sc pseudobulk input (ref.) | 0.967 ± 0.001 | 0.989 ± 0.000 | 0.905 ± 0.001 | 0.0446 ± 0.0003 | 0.1325 ± 0.0008 | 2.256 ± 0.052 | 0.792 ± 0.005 | 0.357 ± 0.003 | 0.871 ± 0.003 | 0.429 ± 0.003 | 0.938 ± 0.001 | 0.531 ± 0.004 |
+| Sepsis | Real vs real control | 0.501 ± 0.001 | 0.495 ± 0.004 | 0.503 ± 0.001 | -3.9e-05 ± 0.0000 | -0.0001 ± 0.0000 | 0.100 ± 0.004 | 0.874 ± 0.002 | 0.925 ± 0.002 | 0.938 ± 0.001 | 0.986 ± 0.001 | 0.981 ± 0.001 | 1.000 ± 0.000 |
+| Sepsis | Zero-shot, real bulk | 0.997 ± 0.000 | 0.999 ± 0.000 | 0.990 ± 0.000 | 0.1914 ± 0.0035 | 0.4280 ± 0.0064 | 5.160 ± 0.240 | 0.318 ± 0.036 | 0.034 ± 0.001 | 0.435 ± 0.030 | 0.048 ± 0.001 | 0.625 ± 0.012 | 0.076 ± 0.001 |
+| Sepsis | Zero-shot + calibration | 0.984 ± 0.001 | 0.988 ± 0.002 | 0.940 ± 0.006 | 0.0404 ± 0.0125 | 0.1242 ± 0.0343 | 1.950 ± 0.150 | 0.603 ± 0.009 | 0.220 ± 0.026 | 0.752 ± 0.010 | 0.298 ± 0.034 | 0.899 ± 0.005 | 0.428 ± 0.046 |
+| Sepsis | scSemiProfiler, 5 donors | 0.998 ± 0.001 | 1.000 ± 0.000 | 0.992 ± 0.002 | 0.1666 ± 0.0297 | 0.4743 ± 0.0940 | 3.830 ± 0.347 | 0.408 ± 0.057 | 0.040 ± 0.007 | 0.541 ± 0.063 | 0.056 ± 0.009 | 0.728 ± 0.062 | 0.080 ± 0.010 |
+| Sepsis | **LoRA, 5 donors** | 0.936 ± 0.010 | 0.942 ± 0.007 | 0.840 ± 0.017 | 0.0210 ± 0.0073 | 0.0606 ± 0.0187 | 1.461 ± 0.121 | 0.649 ± 0.035 | 0.419 ± 0.043 | 0.778 ± 0.029 | 0.536 ± 0.045 | 0.906 ± 0.020 | 0.694 ± 0.042 |
+| Sepsis | Zero-shot, sc pseudobulk input (ref.) | 0.980 ± 0.001 | 0.984 ± 0.001 | 0.929 ± 0.001 | 0.0285 ± 0.0006 | 0.0886 ± 0.0017 | 1.620 ± 0.055 | 0.644 ± 0.008 | 0.266 ± 0.003 | 0.787 ± 0.004 | 0.357 ± 0.006 | 0.918 ± 0.002 | 0.498 ± 0.008 |
+| Imperial | Real vs real control | 0.498 ± 0.003 | 0.501 ± 0.004 | 0.499 ± 0.002 | -1.3e-05 ± 0.0000 | -3.5e-05 ± 0.0000 | 0.260 ± 0.010 | 0.842 ± 0.002 | 0.919 ± 0.002 | 0.915 ± 0.001 | 0.985 ± 0.001 | 0.969 ± 0.000 | 1.000 ± 0.000 |
+| Imperial | Zero-shot, real bulk | 0.995 ± 0.000 | 0.998 ± 0.000 | 0.981 ± 0.001 | 0.2425 ± 0.0042 | 0.5198 ± 0.0078 | 4.277 ± 0.148 | 0.221 ± 0.015 | 0.049 ± 0.001 | 0.314 ± 0.014 | 0.071 ± 0.002 | 0.500 ± 0.016 | 0.114 ± 0.004 |
+| Imperial | Zero-shot + calibration | 0.961 ± 0.003 | 0.967 ± 0.002 | 0.877 ± 0.005 | 0.0218 ± 0.0021 | 0.0673 ± 0.0062 | 1.216 ± 0.023 | 0.598 ± 0.016 | 0.375 ± 0.006 | 0.719 ± 0.014 | 0.486 ± 0.006 | 0.845 ± 0.009 | 0.634 ± 0.006 |
+| Imperial | scSemiProfiler, 5 donors | 0.999 ± 0.000 | 1.000 ± 0.000 | 0.995 ± 0.001 | 0.1372 ± 0.0079 | 0.3741 ± 0.0278 | 2.039 ± 0.074 | 0.117 ± 0.026 | 0.025 ± 0.005 | 0.186 ± 0.029 | 0.037 ± 0.007 | 0.330 ± 0.029 | 0.059 ± 0.011 |
+| Imperial | **LoRA, 5 donors** | 0.918 ± 0.011 | 0.916 ± 0.012 | 0.804 ± 0.016 | 0.0156 ± 0.0052 | 0.0443 ± 0.0147 | 1.282 ± 0.159 | 0.610 ± 0.030 | 0.470 ± 0.064 | 0.726 ± 0.028 | 0.589 ± 0.079 | 0.853 ± 0.018 | 0.737 ± 0.085 |
+| Imperial | Zero-shot, sc pseudobulk input (ref.) | 0.959 ± 0.001 | 0.966 ± 0.001 | 0.870 ± 0.001 | 0.0174 ± 0.0002 | 0.0503 ± 0.0005 | 1.105 ± 0.018 | 0.583 ± 0.004 | 0.419 ± 0.006 | 0.705 ± 0.006 | 0.541 ± 0.006 | 0.840 ± 0.005 | 0.703 ± 0.007 |
+| Giroux | Real vs real control | 0.495 ± 0.005 | 0.496 ± 0.016 | 0.506 ± 0.011 | -1.9e-05 ± 0.0001 | -3.3e-06 ± 0.0003 | 0.145 ± 0.034 | 0.991 ± 0.003 | 0.993 ± 0.004 | 0.997 ± 0.002 | 0.999 ± 0.001 | 0.999 ± 0.001 | 1.000 ± 0.000 |
+| Giroux | Zero-shot, real bulk | 0.993 ± 0.002 | 0.996 ± 0.001 | 0.971 ± 0.004 | 0.1039 ± 0.0089 | 0.2867 ± 0.0200 | 5.264 ± 0.874 | 0.860 ± 0.014 | 0.300 ± 0.043 | 0.926 ± 0.009 | 0.363 ± 0.053 | 0.980 ± 0.009 | 0.472 ± 0.068 |
+| Giroux | Zero-shot + calibration | 0.942 ± 0.007 | 0.947 ± 0.005 | 0.861 ± 0.011 | 0.0257 ± 0.0042 | 0.0838 ± 0.0117 | 2.716 ± 0.587 | 0.958 ± 0.019 | 0.780 ± 0.025 | 0.981 ± 0.010 | 0.864 ± 0.021 | 0.996 ± 0.004 | 0.944 ± 0.014 |
+| Giroux | scSemiProfiler, 5 donors | 0.997 ± 0.001 | 0.999 ± 0.001 | 0.987 ± 0.002 | 0.0743 ± 0.0084 | 0.2210 ± 0.0255 | 3.795 ± 0.573 | 0.881 ± 0.016 | 0.373 ± 0.024 | 0.951 ± 0.011 | 0.468 ± 0.025 | 0.992 ± 0.004 | 0.607 ± 0.025 |
+| Giroux | **LoRA, 5 donors** | 0.893 ± 0.009 | 0.905 ± 0.010 | 0.793 ± 0.016 | 0.0178 ± 0.0053 | 0.0584 ± 0.0155 | 1.990 ± 0.413 | 0.983 ± 0.004 | 0.830 ± 0.036 | 0.995 ± 0.001 | 0.903 ± 0.022 | 1.000 ± 0.000 | 0.961 ± 0.010 |
+| Giroux | Zero-shot, sc pseudobulk input (ref.) | 0.931 ± 0.002 | 0.940 ± 0.004 | 0.839 ± 0.007 | 0.0215 ± 0.0020 | 0.0738 ± 0.0067 | 2.219 ± 0.371 | 0.974 ± 0.013 | 0.770 ± 0.014 | 0.988 ± 0.008 | 0.849 ± 0.011 | 0.997 ± 0.003 | 0.928 ± 0.008 |
+| Ji (whole blood) | Real vs real control | 0.491 ± 0.006 | 0.489 ± 0.011 | 0.495 ± 0.005 | -2.8e-06 ± 0.0001 | -2.6e-05 ± 0.0002 | 0.073 ± 0.023 | 0.972 ± 0.005 | 0.980 ± 0.004 | 0.989 ± 0.003 | 0.998 ± 0.001 | 0.997 ± 0.001 | 1.000 ± 0.000 |
+| Ji (whole blood) | Zero-shot, real bulk | 0.999 ± 0.001 | 1.000 ± 0.000 | 0.996 ± 0.001 | 0.2734 ± 0.0038 | 0.6915 ± 0.0081 | 32.986 ± 15.866 | 0.412 ± 0.039 | 0.031 ± 0.001 | 0.558 ± 0.077 | 0.042 ± 0.002 | 0.799 ± 0.015 | 0.066 ± 0.005 |
+| Ji (whole blood) | Zero-shot + calibration | 0.998 ± 0.000 | 0.999 ± 0.000 | 0.992 ± 0.001 | 0.1382 ± 0.0083 | 0.4002 ± 0.0218 | 10.389 ± 2.366 | 0.409 ± 0.031 | 0.057 ± 0.008 | 0.556 ± 0.051 | 0.076 ± 0.011 | 0.782 ± 0.027 | 0.114 ± 0.014 |
+| Ji (whole blood) | scSemiProfiler, 5 donors | 0.996 ± 0.001 | 0.999 ± 0.000 | 0.991 ± 0.003 | 0.1527 ± 0.0078 | 0.4520 ± 0.0267 | 4.268 ± 0.468 | 0.557 ± 0.074 | 0.081 ± 0.016 | 0.685 ± 0.076 | 0.105 ± 0.019 | 0.840 ± 0.054 | 0.145 ± 0.021 |
+| Ji (whole blood) | **LoRA, 5 donors** | 0.955 ± 0.015 | 0.956 ± 0.011 | 0.874 ± 0.026 | 0.0408 ± 0.0211 | 0.1276 ± 0.0645 | 2.604 ± 1.098 | 0.849 ± 0.054 | 0.478 ± 0.116 | 0.922 ± 0.034 | 0.579 ± 0.126 | 0.978 ± 0.010 | 0.707 ± 0.119 |
+| Ji (whole blood) | Zero-shot, sc pseudobulk input (ref.) | 0.962 ± 0.003 | 0.980 ± 0.003 | 0.896 ± 0.006 | 0.0256 ± 0.0009 | 0.0848 ± 0.0029 | 1.656 ± 0.146 | 0.910 ± 0.018 | 0.494 ± 0.014 | 0.958 ± 0.008 | 0.591 ± 0.018 | 0.988 ± 0.003 | 0.718 ± 0.021 |
+
+### (a) 5-donor few-pair design: mean ± SD over the 5 draws: is LoRA better than each comparator? (draws where LoRA is better / draws)
+
+| cohort | LoRA vs | AUC kNN | AUC RF | 1-NN acc | MMD² (σ=m) | MMD² (σ sum) | E_N (all cells) | P k3 | C k3 | P k5 | C k5 | P k10 | C k10 | E_N (P/C cells) |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| ImMicro | scSemiProfiler, 5 donors | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 |
+| ImMicro | Zero-shot, real bulk | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 |
+| ImMicro | Zero-shot + calibration | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 |
+| Sepsis | scSemiProfiler, 5 donors | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 |
+| Sepsis | Zero-shot, real bulk | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 |
+| Sepsis | Zero-shot + calibration | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 4/5 | 5/5 | 3/5 | 5/5 | 5/5 |
+| Imperial | scSemiProfiler, 5 donors | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 |
+| Imperial | Zero-shot, real bulk | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 |
+| Imperial | Zero-shot + calibration | 5/5 | 5/5 | 5/5 | 4/5 | 4/5 | 2/5 | 3/5 | 5/5 | 3/5 | 5/5 | 3/5 | 5/5 | 2/5 |
+| Giroux | scSemiProfiler, 5 donors | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 0/5 |
+| Giroux | Zero-shot, real bulk | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 |
+| Giroux | Zero-shot + calibration | 5/5 | 5/5 | 5/5 | 4/5 | 4/5 | 5/5 | 5/5 | 5/5 | 4/5 | 5/5 | 4/5 | 5/5 | 5/5 |
+| Ji (whole blood) | scSemiProfiler, 5 donors | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 4/5 |
+| Ji (whole blood) | Zero-shot, real bulk | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 |
+| Ji (whole blood) | Zero-shot + calibration | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 |
+
+Exceptions (LoRA worse; ties are not counted as worse):
+
+- Sepsis: LoRA worse than Zero-shot + calibration on P k5 in 1/5 draws
+- Sepsis: LoRA worse than Zero-shot + calibration on P k10 in 2/5 draws
+- Imperial: LoRA worse than Zero-shot + calibration on MMD² (σ=m) in 1/5 draws
+- Imperial: LoRA worse than Zero-shot + calibration on MMD² (σ sum) in 1/5 draws
+- Imperial: LoRA worse than Zero-shot + calibration on E_N (all cells) in 3/5 draws
+- Imperial: LoRA worse than Zero-shot + calibration on P k3 in 2/5 draws
+- Imperial: LoRA worse than Zero-shot + calibration on P k5 in 2/5 draws
+- Imperial: LoRA worse than Zero-shot + calibration on P k10 in 2/5 draws
+- Imperial: LoRA worse than Zero-shot + calibration on E_N (P/C cells) in 3/5 draws
+- Giroux: LoRA worse than scSemiProfiler, 5 donors on E_N (P/C cells) in 5/5 draws
+- Giroux: LoRA worse than Zero-shot + calibration on MMD² (σ=m) in 1/5 draws
+- Giroux: LoRA worse than Zero-shot + calibration on MMD² (σ sum) in 1/5 draws
+- Giroux: LoRA worse than Zero-shot + calibration on P k5 in 1/5 draws
+- Giroux: LoRA worse than Zero-shot + calibration on P k10 in 1/5 draws
+- Ji (whole blood): LoRA worse than scSemiProfiler, 5 donors on E_N (P/C cells) in 1/5 draws
+
+### (a) 5-donor few-pair design: mean ± SD over the 5 draws: Kendall τ of each metric's arm ranking with the coverage / precision (k = 10) ranking (arms only; few-pair: means over draws, E_N medians)
+
+| cohort | vs | AUC kNN | AUC RF | 1-NN acc | MMD² (σ=m) | MMD² (σ sum) | E_N (all cells) | P k3 | C k3 | P k5 | C k5 | P k10 | C k10 | E_N (P/C cells) |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| ImMicro | coverage k10 | 0.80 | 1.00 | 0.80 | 1.00 | 1.00 | 0.60 | 0.80 | 0.80 | 0.80 | 0.80 | 0.80 | 1.00 | 0.40 |
+| ImMicro | precision k10 | 1.00 | 0.80 | 1.00 | 0.80 | 0.80 | 0.80 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 0.80 | 0.60 |
+| Sepsis | coverage k10 | 0.80 | 0.80 | 0.80 | 1.00 | 0.80 | 1.00 | 1.00 | 1.00 | 0.80 | 1.00 | 0.80 | 1.00 | 0.80 |
+| Sepsis | precision k10 | 0.60 | 0.60 | 0.60 | 0.80 | 0.60 | 0.80 | 0.80 | 0.80 | 1.00 | 0.80 | 1.00 | 0.80 | 1.00 |
+| Imperial | coverage k10 | 1.00 | 1.00 | 1.00 | 0.80 | 0.80 | 0.40 | 0.80 | 1.00 | 0.80 | 1.00 | 0.80 | 1.00 | 0.40 |
+| Imperial | precision k10 | 0.80 | 0.80 | 0.80 | 0.60 | 0.60 | 0.20 | 1.00 | 0.80 | 1.00 | 0.80 | 1.00 | 0.80 | 0.20 |
+| Giroux | coverage k10 | 0.60 | 0.60 | 0.60 | 0.80 | 0.80 | 0.80 | 0.80 | 1.00 | 0.80 | 1.00 | 0.80 | 1.00 | 0.20 |
+| Giroux | precision k10 | 0.80 | 0.80 | 0.80 | 1.00 | 1.00 | 1.00 | 1.00 | 0.80 | 1.00 | 0.80 | 1.00 | 0.80 | 0.40 |
+| Ji (whole blood) | coverage k10 | 0.80 | 0.80 | 0.80 | 0.80 | 0.80 | 1.00 | 0.80 | 1.00 | 0.80 | 1.00 | 0.80 | 1.00 | 1.00 |
+| Ji (whole blood) | precision k10 | 0.60 | 0.60 | 0.60 | 0.60 | 0.60 | 0.80 | 1.00 | 0.80 | 1.00 | 0.80 | 1.00 | 0.80 | 0.80 |
+
+### (a) 5-donor few-pair design: mean ± SD over the 5 draws: every variant (real side R1 / R2, raw / depth matched)
+
+#### real side R1, no depth matching
+
+| cohort | set | AUC kNN | AUC RF | 1-NN acc | MMD² (σ=m) | MMD² (σ sum) | P k3 | C k3 | P k5 | C k5 | P k10 | C k10 | E_N (all cells) | E_N (P/C cells) |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| ImMicro | Real vs real control (R1 vs R2) | 0.507 ± 0.002 | 0.584 ± 0.002 | 0.499 ± 0.001 | 3.5e-05 ± 0.0000 | 0.0002 ± 0.0000 | 0.835 ± 0.001 | 0.932 ± 0.002 | 0.907 ± 0.001 | 0.988 ± 0.001 | 0.965 ± 0.001 | 1.000 ± 0.000 | 0.382 ± 0.015 | 0.484 ± 0.008 |
+| ImMicro | Zero-shot, real bulk | 0.995 ± 0.000 | 0.998 ± 0.000 | 0.978 ± 0.001 | 0.1003 ± 0.0006 | 0.2400 ± 0.0013 | 0.399 ± 0.009 | 0.087 ± 0.003 | 0.521 ± 0.003 | 0.122 ± 0.003 | 0.677 ± 0.006 | 0.190 ± 0.004 | 7.236 ± 0.345 | 34.644 ± 5.613 |
+| ImMicro | Zero-shot + calibration | 0.966 ± 0.002 | 0.986 ± 0.001 | 0.896 ± 0.002 | 0.0429 ± 0.0025 | 0.1201 ± 0.0069 | 0.706 ± 0.014 | 0.346 ± 0.003 | 0.807 ± 0.008 | 0.425 ± 0.005 | 0.904 ± 0.006 | 0.535 ± 0.008 | 2.855 ± 0.085 | 2.885 ± 0.123 |
+| ImMicro | scSemiProfiler, 5 donors | 0.998 ± 0.000 | 1.000 ± 0.000 | 0.995 ± 0.000 | 0.1851 ± 0.0092 | 0.4925 ± 0.0241 | 0.236 ± 0.025 | 0.019 ± 0.001 | 0.333 ± 0.034 | 0.025 ± 0.001 | 0.518 ± 0.036 | 0.036 ± 0.002 | 3.649 ± 0.046 | 2.747 ± 0.049 |
+| ImMicro | LoRA, 5 donors | 0.935 ± 0.004 | 0.964 ± 0.002 | 0.823 ± 0.010 | 0.0231 ± 0.0046 | 0.0689 ± 0.0129 | 0.797 ± 0.017 | 0.564 ± 0.039 | 0.882 ± 0.012 | 0.671 ± 0.045 | 0.952 ± 0.005 | 0.792 ± 0.047 | 1.547 ± 0.076 | 1.452 ± 0.110 |
+| ImMicro | Zero-shot, sc pseudobulk input (ref.) | 0.959 ± 0.001 | 0.989 ± 0.000 | 0.890 ± 0.001 | 0.0448 ± 0.0004 | 0.1345 ± 0.0011 | 0.792 ± 0.005 | 0.357 ± 0.003 | 0.871 ± 0.003 | 0.429 ± 0.003 | 0.938 ± 0.001 | 0.531 ± 0.004 | 2.256 ± 0.052 | 2.034 ± 0.040 |
+| Sepsis | Real vs real control (R1 vs R2) | 0.517 ± 0.001 | 0.596 ± 0.002 | 0.507 ± 0.001 | 2.2e-05 ± 0.0000 | 0.0001 ± 0.0000 | 0.874 ± 0.002 | 0.925 ± 0.002 | 0.938 ± 0.001 | 0.986 ± 0.001 | 0.981 ± 0.001 | 1.000 ± 0.000 | 0.100 ± 0.004 | 0.207 ± 0.006 |
+| Sepsis | Zero-shot, real bulk | 0.997 ± 0.000 | 0.999 ± 0.000 | 0.989 ± 0.000 | 0.1889 ± 0.0037 | 0.4189 ± 0.0069 | 0.318 ± 0.036 | 0.034 ± 0.001 | 0.435 ± 0.030 | 0.048 ± 0.001 | 0.625 ± 0.012 | 0.076 ± 0.001 | 5.160 ± 0.240 | 16.070 ± 2.555 |
+| Sepsis | Zero-shot + calibration | 0.982 ± 0.001 | 0.987 ± 0.002 | 0.936 ± 0.006 | 0.0407 ± 0.0123 | 0.1256 ± 0.0332 | 0.603 ± 0.009 | 0.220 ± 0.026 | 0.752 ± 0.010 | 0.298 ± 0.034 | 0.899 ± 0.005 | 0.428 ± 0.046 | 1.950 ± 0.150 | 1.413 ± 0.108 |
+| Sepsis | scSemiProfiler, 5 donors | 0.997 ± 0.001 | 1.000 ± 0.000 | 0.990 ± 0.002 | 0.1658 ± 0.0297 | 0.4712 ± 0.0943 | 0.408 ± 0.057 | 0.040 ± 0.007 | 0.541 ± 0.063 | 0.056 ± 0.009 | 0.728 ± 0.062 | 0.080 ± 0.010 | 3.830 ± 0.347 | 2.146 ± 0.138 |
+| Sepsis | LoRA, 5 donors | 0.936 ± 0.009 | 0.944 ± 0.006 | 0.843 ± 0.015 | 0.0214 ± 0.0070 | 0.0626 ± 0.0177 | 0.649 ± 0.035 | 0.419 ± 0.043 | 0.778 ± 0.029 | 0.536 ± 0.045 | 0.906 ± 0.020 | 0.694 ± 0.042 | 1.461 ± 0.121 | 1.225 ± 0.147 |
+| Sepsis | Zero-shot, sc pseudobulk input (ref.) | 0.977 ± 0.001 | 0.983 ± 0.001 | 0.921 ± 0.001 | 0.0290 ± 0.0005 | 0.0906 ± 0.0016 | 0.644 ± 0.008 | 0.266 ± 0.003 | 0.787 ± 0.004 | 0.357 ± 0.006 | 0.918 ± 0.002 | 0.498 ± 0.008 | 1.620 ± 0.055 | 1.128 ± 0.038 |
+| Imperial | Real vs real control (R1 vs R2) | 0.502 ± 0.001 | 0.560 ± 0.001 | 0.503 ± 0.001 | -2.9e-06 ± 0.0000 | 4.4e-05 ± 0.0000 | 0.842 ± 0.002 | 0.919 ± 0.002 | 0.915 ± 0.001 | 0.985 ± 0.001 | 0.969 ± 0.000 | 1.000 ± 0.000 | 0.260 ± 0.010 | 0.233 ± 0.006 |
+| Imperial | Zero-shot, real bulk | 0.995 ± 0.000 | 0.997 ± 0.000 | 0.980 ± 0.001 | 0.2410 ± 0.0039 | 0.5131 ± 0.0073 | 0.221 ± 0.015 | 0.049 ± 0.001 | 0.314 ± 0.014 | 0.071 ± 0.002 | 0.500 ± 0.016 | 0.114 ± 0.004 | 4.277 ± 0.148 | 6.844 ± 0.418 |
+| Imperial | Zero-shot + calibration | 0.958 ± 0.003 | 0.967 ± 0.002 | 0.872 ± 0.004 | 0.0220 ± 0.0021 | 0.0684 ± 0.0063 | 0.598 ± 0.016 | 0.375 ± 0.006 | 0.719 ± 0.014 | 0.486 ± 0.006 | 0.845 ± 0.009 | 0.634 ± 0.006 | 1.216 ± 0.023 | 1.167 ± 0.029 |
+| Imperial | scSemiProfiler, 5 donors | 0.999 ± 0.000 | 1.000 ± 0.000 | 0.994 ± 0.002 | 0.1373 ± 0.0079 | 0.3732 ± 0.0278 | 0.117 ± 0.026 | 0.025 ± 0.005 | 0.186 ± 0.029 | 0.037 ± 0.007 | 0.330 ± 0.029 | 0.059 ± 0.011 | 2.039 ± 0.074 | 1.683 ± 0.074 |
+| Imperial | LoRA, 5 donors | 0.914 ± 0.012 | 0.918 ± 0.012 | 0.804 ± 0.018 | 0.0157 ± 0.0051 | 0.0448 ± 0.0145 | 0.610 ± 0.030 | 0.470 ± 0.064 | 0.726 ± 0.028 | 0.589 ± 0.079 | 0.853 ± 0.018 | 0.737 ± 0.085 | 1.282 ± 0.159 | 1.284 ± 0.216 |
+| Imperial | Zero-shot, sc pseudobulk input (ref.) | 0.956 ± 0.001 | 0.966 ± 0.001 | 0.863 ± 0.002 | 0.0175 ± 0.0002 | 0.0512 ± 0.0005 | 0.583 ± 0.004 | 0.419 ± 0.006 | 0.705 ± 0.006 | 0.541 ± 0.006 | 0.840 ± 0.005 | 0.703 ± 0.007 | 1.105 ± 0.018 | 1.032 ± 0.031 |
+| Giroux | Real vs real control (R1 vs R2) | 0.589 ± 0.009 | 0.935 ± 0.008 | 0.533 ± 0.005 | 0.0021 ± 0.0003 | 0.0138 ± 0.0017 | 0.991 ± 0.003 | 0.993 ± 0.004 | 0.997 ± 0.002 | 0.999 ± 0.001 | 0.999 ± 0.001 | 1.000 ± 0.000 | 0.145 ± 0.034 | 0.287 ± 0.028 |
+| Giroux | Zero-shot, real bulk | 0.960 ± 0.008 | 0.996 ± 0.001 | 0.907 ± 0.013 | 0.1003 ± 0.0085 | 0.2666 ± 0.0181 | 0.860 ± 0.014 | 0.300 ± 0.043 | 0.926 ± 0.009 | 0.363 ± 0.053 | 0.980 ± 0.009 | 0.472 ± 0.068 | 5.264 ± 0.874 | 8.853 ± 2.634 |
+| Giroux | Zero-shot + calibration | 0.865 ± 0.017 | 0.977 ± 0.001 | 0.731 ± 0.018 | 0.0269 ± 0.0038 | 0.0928 ± 0.0096 | 0.958 ± 0.019 | 0.780 ± 0.025 | 0.981 ± 0.010 | 0.864 ± 0.021 | 0.996 ± 0.004 | 0.944 ± 0.014 | 2.716 ± 0.587 | 1.814 ± 0.401 |
+| Giroux | scSemiProfiler, 5 donors | 0.968 ± 0.002 | 0.999 ± 0.000 | 0.908 ± 0.008 | 0.0743 ± 0.0074 | 0.2180 ± 0.0200 | 0.881 ± 0.016 | 0.373 ± 0.024 | 0.951 ± 0.011 | 0.468 ± 0.025 | 0.992 ± 0.004 | 0.607 ± 0.025 | 3.795 ± 0.573 | 1.206 ± 0.039 |
+| Giroux | LoRA, 5 donors | 0.818 ± 0.020 | 0.971 ± 0.003 | 0.694 ± 0.024 | 0.0228 ± 0.0052 | 0.0884 ± 0.0147 | 0.983 ± 0.004 | 0.830 ± 0.036 | 0.995 ± 0.001 | 0.903 ± 0.022 | 1.000 ± 0.000 | 0.961 ± 0.010 | 1.990 ± 0.413 | 1.450 ± 0.217 |
+| Giroux | Zero-shot, sc pseudobulk input (ref.) | 0.851 ± 0.012 | 0.975 ± 0.002 | 0.725 ± 0.008 | 0.0247 ± 0.0016 | 0.0956 ± 0.0057 | 0.974 ± 0.013 | 0.770 ± 0.014 | 0.988 ± 0.008 | 0.849 ± 0.011 | 0.997 ± 0.003 | 0.928 ± 0.008 | 2.219 ± 0.371 | 1.164 ± 0.103 |
+| Ji (whole blood) | Real vs real control (R1 vs R2) | 0.557 ± 0.007 | 0.810 ± 0.005 | 0.526 ± 0.009 | 0.0002 ± 0.0000 | 0.0017 ± 0.0002 | 0.972 ± 0.005 | 0.980 ± 0.004 | 0.989 ± 0.003 | 0.998 ± 0.001 | 0.997 ± 0.001 | 1.000 ± 0.000 | 0.073 ± 0.023 | 0.209 ± 0.022 |
+| Ji (whole blood) | Zero-shot, real bulk | 0.997 ± 0.001 | 1.000 ± 0.000 | 0.991 ± 0.001 | 0.2708 ± 0.0022 | 0.6730 ± 0.0073 | 0.412 ± 0.039 | 0.031 ± 0.001 | 0.558 ± 0.077 | 0.042 ± 0.002 | 0.799 ± 0.015 | 0.066 ± 0.005 | 32.986 ± 15.866 | — †5 |
+| Ji (whole blood) | Zero-shot + calibration | 0.996 ± 0.001 | 0.999 ± 0.000 | 0.985 ± 0.002 | 0.1384 ± 0.0083 | 0.3948 ± 0.0212 | 0.409 ± 0.031 | 0.057 ± 0.008 | 0.556 ± 0.051 | 0.076 ± 0.011 | 0.782 ± 0.027 | 0.114 ± 0.014 | 10.389 ± 2.366 | 13.620 ± 2.630 †1 |
+| Ji (whole blood) | scSemiProfiler, 5 donors | 0.992 ± 0.001 | 0.999 ± 0.000 | 0.980 ± 0.005 | 0.1520 ± 0.0086 | 0.4460 ± 0.0291 | 0.557 ± 0.074 | 0.081 ± 0.016 | 0.685 ± 0.076 | 0.105 ± 0.019 | 0.840 ± 0.054 | 0.145 ± 0.021 | 4.268 ± 0.468 | 2.205 ± 0.118 |
+| Ji (whole blood) | LoRA, 5 donors | 0.939 ± 0.020 | 0.969 ± 0.006 | 0.846 ± 0.039 | 0.0423 ± 0.0213 | 0.1338 ± 0.0651 | 0.849 ± 0.054 | 0.478 ± 0.116 | 0.922 ± 0.034 | 0.579 ± 0.126 | 0.978 ± 0.010 | 0.707 ± 0.119 | 2.604 ± 1.098 | 1.985 ± 0.563 |
+| Ji (whole blood) | Zero-shot, sc pseudobulk input (ref.) | 0.936 ± 0.005 | 0.984 ± 0.003 | 0.848 ± 0.008 | 0.0271 ± 0.0009 | 0.0921 ± 0.0025 | 0.910 ± 0.018 | 0.494 ± 0.014 | 0.958 ± 0.008 | 0.591 ± 0.018 | 0.988 ± 0.003 | 0.718 ± 0.021 | 1.656 ± 0.146 | 1.100 ± 0.081 |
+
+#### real side R1, depth matched
+
+| cohort | set | AUC kNN | AUC RF | 1-NN acc | MMD² (σ=m) | MMD² (σ sum) | P k3 | C k3 | P k5 | C k5 | P k10 | C k10 |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| ImMicro | Real vs real control (R1 vs R2) | 0.507 ± 0.003 | 0.581 ± 0.003 | 0.499 ± 0.002 | 3.2e-05 ± 0.0000 | 0.0002 ± 0.0000 | 0.833 ± 0.001 | 0.931 ± 0.002 | 0.905 ± 0.001 | 0.987 ± 0.001 | 0.964 ± 0.001 | 1.000 ± 0.000 |
+| ImMicro | Zero-shot, real bulk | 0.995 ± 0.000 | 0.998 ± 0.000 | 0.975 ± 0.000 | 0.0948 ± 0.0006 | 0.2250 ± 0.0013 | 0.394 ± 0.011 | 0.097 ± 0.002 | 0.512 ± 0.014 | 0.137 ± 0.003 | 0.671 ± 0.010 | 0.213 ± 0.003 |
+| ImMicro | Zero-shot + calibration | 0.933 ± 0.004 | 0.976 ± 0.001 | 0.828 ± 0.003 | 0.0271 ± 0.0011 | 0.0805 ± 0.0032 | 0.801 ± 0.009 | 0.552 ± 0.007 | 0.875 ± 0.006 | 0.657 ± 0.006 | 0.940 ± 0.004 | 0.780 ± 0.006 |
+| ImMicro | scSemiProfiler, 5 donors | 0.987 ± 0.002 | 0.999 ± 0.000 | 0.962 ± 0.004 | 0.0912 ± 0.0045 | 0.2568 ± 0.0115 | 0.631 ± 0.012 | 0.151 ± 0.017 | 0.728 ± 0.007 | 0.200 ± 0.020 | 0.829 ± 0.007 | 0.284 ± 0.028 |
+| ImMicro | LoRA, 5 donors | 0.870 ± 0.006 | 0.943 ± 0.003 | 0.721 ± 0.005 | 0.0118 ± 0.0018 | 0.0383 ± 0.0055 | 0.888 ± 0.007 | 0.813 ± 0.011 | 0.936 ± 0.005 | 0.900 ± 0.008 | 0.975 ± 0.002 | 0.966 ± 0.006 |
+| ImMicro | Zero-shot, sc pseudobulk input (ref.) | 0.927 ± 0.001 | 0.982 ± 0.000 | 0.831 ± 0.001 | 0.0304 ± 0.0003 | 0.0974 ± 0.0007 | 0.867 ± 0.003 | 0.537 ± 0.003 | 0.921 ± 0.001 | 0.633 ± 0.004 | 0.965 ± 0.001 | 0.751 ± 0.003 |
+| Sepsis | Real vs real control (R1 vs R2) | 0.513 ± 0.002 | 0.592 ± 0.002 | 0.507 ± 0.002 | 2.2e-05 ± 0.0000 | 0.0001 ± 0.0000 | 0.870 ± 0.003 | 0.924 ± 0.002 | 0.937 ± 0.002 | 0.987 ± 0.001 | 0.981 ± 0.001 | 1.000 ± 0.000 |
+| Sepsis | Zero-shot, real bulk | 0.997 ± 0.000 | 0.999 ± 0.000 | 0.989 ± 0.000 | 0.1907 ± 0.0034 | 0.4261 ± 0.0062 | 0.305 ± 0.023 | 0.032 ± 0.001 | 0.417 ± 0.014 | 0.045 ± 0.000 | 0.595 ± 0.012 | 0.073 ± 0.001 |
+| Sepsis | Zero-shot + calibration | 0.960 ± 0.005 | 0.973 ± 0.004 | 0.879 ± 0.012 | 0.0318 ± 0.0117 | 0.0956 ± 0.0322 | 0.658 ± 0.009 | 0.390 ± 0.036 | 0.784 ± 0.010 | 0.503 ± 0.042 | 0.909 ± 0.007 | 0.658 ± 0.048 |
+| Sepsis | scSemiProfiler, 5 donors | 0.995 ± 0.001 | 0.999 ± 0.000 | 0.983 ± 0.004 | 0.1062 ± 0.0256 | 0.3114 ± 0.0773 | 0.261 ± 0.063 | 0.069 ± 0.014 | 0.360 ± 0.069 | 0.099 ± 0.019 | 0.504 ± 0.073 | 0.158 ± 0.028 |
+| Sepsis | LoRA, 5 donors | 0.890 ± 0.012 | 0.917 ± 0.011 | 0.783 ± 0.012 | 0.0163 ± 0.0076 | 0.0468 ± 0.0193 | 0.700 ± 0.045 | 0.567 ± 0.023 | 0.808 ± 0.038 | 0.694 ± 0.024 | 0.914 ± 0.025 | 0.834 ± 0.023 |
+| Sepsis | Zero-shot, sc pseudobulk input (ref.) | 0.946 ± 0.001 | 0.965 ± 0.000 | 0.856 ± 0.002 | 0.0195 ± 0.0002 | 0.0596 ± 0.0007 | 0.725 ± 0.009 | 0.468 ± 0.006 | 0.835 ± 0.006 | 0.595 ± 0.006 | 0.936 ± 0.002 | 0.753 ± 0.007 |
+| Imperial | Real vs real control (R1 vs R2) | 0.502 ± 0.001 | 0.560 ± 0.002 | 0.503 ± 0.001 | -5.1e-06 ± 0.0000 | 3.4e-05 ± 0.0000 | 0.841 ± 0.002 | 0.919 ± 0.002 | 0.914 ± 0.001 | 0.985 ± 0.001 | 0.969 ± 0.001 | 1.000 ± 0.000 |
+| Imperial | Zero-shot, real bulk | 0.994 ± 0.000 | 0.996 ± 0.000 | 0.973 ± 0.001 | 0.2225 ± 0.0041 | 0.4702 ± 0.0080 | 0.246 ± 0.022 | 0.068 ± 0.003 | 0.359 ± 0.030 | 0.097 ± 0.003 | 0.561 ± 0.028 | 0.151 ± 0.006 |
+| Imperial | Zero-shot + calibration | 0.950 ± 0.003 | 0.966 ± 0.002 | 0.859 ± 0.004 | 0.0237 ± 0.0019 | 0.0723 ± 0.0060 | 0.650 ± 0.014 | 0.413 ± 0.008 | 0.757 ± 0.012 | 0.527 ± 0.010 | 0.868 ± 0.009 | 0.681 ± 0.009 |
+| Imperial | scSemiProfiler, 5 donors | 0.997 ± 0.001 | 0.999 ± 0.000 | 0.986 ± 0.002 | 0.1173 ± 0.0125 | 0.3286 ± 0.0361 | 0.202 ± 0.025 | 0.063 ± 0.009 | 0.290 ± 0.031 | 0.093 ± 0.014 | 0.429 ± 0.044 | 0.152 ± 0.022 |
+| Imperial | LoRA, 5 donors | 0.880 ± 0.012 | 0.903 ± 0.007 | 0.761 ± 0.012 | 0.0100 ± 0.0021 | 0.0291 ± 0.0065 | 0.659 ± 0.035 | 0.584 ± 0.050 | 0.767 ± 0.030 | 0.720 ± 0.053 | 0.881 ± 0.020 | 0.863 ± 0.044 |
+| Imperial | Zero-shot, sc pseudobulk input (ref.) | 0.946 ± 0.001 | 0.967 ± 0.001 | 0.850 ± 0.003 | 0.0190 ± 0.0003 | 0.0555 ± 0.0007 | 0.662 ± 0.005 | 0.455 ± 0.004 | 0.771 ± 0.004 | 0.576 ± 0.003 | 0.883 ± 0.004 | 0.734 ± 0.004 |
+| Giroux | Real vs real control (R1 vs R2) | 0.592 ± 0.005 | 0.932 ± 0.008 | 0.534 ± 0.004 | 0.0020 ± 0.0003 | 0.0134 ± 0.0017 | 0.991 ± 0.003 | 0.993 ± 0.004 | 0.997 ± 0.001 | 0.999 ± 0.001 | 0.999 ± 0.001 | 1.000 ± 0.000 |
+| Giroux | Zero-shot, real bulk | 0.926 ± 0.017 | 0.997 ± 0.001 | 0.878 ± 0.023 | 0.0723 ± 0.0131 | 0.2037 ± 0.0316 | 0.936 ± 0.014 | 0.357 ± 0.071 | 0.974 ± 0.008 | 0.428 ± 0.085 | 0.996 ± 0.002 | 0.549 ± 0.095 |
+| Giroux | Zero-shot + calibration | 0.875 ± 0.020 | 0.961 ± 0.006 | 0.737 ± 0.019 | 0.0230 ± 0.0027 | 0.0759 ± 0.0075 | 0.928 ± 0.019 | 0.768 ± 0.030 | 0.961 ± 0.013 | 0.857 ± 0.031 | 0.989 ± 0.006 | 0.937 ± 0.024 |
+| Giroux | scSemiProfiler, 5 donors | 0.995 ± 0.001 | 0.999 ± 0.000 | 0.975 ± 0.002 | 0.0925 ± 0.0083 | 0.2796 ± 0.0224 | 0.220 ± 0.025 | 0.108 ± 0.006 | 0.300 ± 0.036 | 0.155 ± 0.006 | 0.426 ± 0.042 | 0.238 ± 0.009 |
+| Giroux | LoRA, 5 donors | 0.795 ± 0.029 | 0.925 ± 0.006 | 0.680 ± 0.028 | 0.0089 ± 0.0013 | 0.0338 ± 0.0043 | 0.942 ± 0.014 | 0.850 ± 0.045 | 0.977 ± 0.006 | 0.923 ± 0.028 | 0.995 ± 0.001 | 0.974 ± 0.011 |
+| Giroux | Zero-shot, sc pseudobulk input (ref.) | 0.839 ± 0.013 | 0.935 ± 0.007 | 0.702 ± 0.009 | 0.0190 ± 0.0027 | 0.0642 ± 0.0085 | 0.939 ± 0.019 | 0.812 ± 0.008 | 0.969 ± 0.013 | 0.892 ± 0.012 | 0.989 ± 0.007 | 0.953 ± 0.009 |
+| Ji (whole blood) | Real vs real control (R1 vs R2) | 0.555 ± 0.004 | 0.806 ± 0.007 | 0.525 ± 0.006 | 0.0002 ± 0.0000 | 0.0016 ± 0.0002 | 0.971 ± 0.004 | 0.980 ± 0.004 | 0.988 ± 0.003 | 0.998 ± 0.001 | 0.997 ± 0.001 | 1.000 ± 0.000 |
+| Ji (whole blood) | Zero-shot, real bulk | 0.996 ± 0.001 | 0.997 ± 0.001 | 0.976 ± 0.003 | 0.1389 ± 0.0021 | 0.3778 ± 0.0066 | 0.223 ± 0.027 | 0.084 ± 0.007 | 0.318 ± 0.040 | 0.121 ± 0.010 | 0.480 ± 0.056 | 0.199 ± 0.017 |
+| Ji (whole blood) | Zero-shot + calibration | 0.992 ± 0.002 | 0.994 ± 0.002 | 0.956 ± 0.006 | 0.0422 ± 0.0045 | 0.1403 ± 0.0126 | 0.342 ± 0.038 | 0.160 ± 0.013 | 0.467 ± 0.043 | 0.230 ± 0.016 | 0.658 ± 0.046 | 0.357 ± 0.025 |
+| Ji (whole blood) | scSemiProfiler, 5 donors | 0.995 ± 0.002 | 0.999 ± 0.000 | 0.986 ± 0.004 | 0.1163 ± 0.0094 | 0.3581 ± 0.0279 | 0.290 ± 0.061 | 0.065 ± 0.014 | 0.398 ± 0.078 | 0.094 ± 0.019 | 0.564 ± 0.098 | 0.150 ± 0.026 |
+| Ji (whole blood) | LoRA, 5 donors | 0.895 ± 0.011 | 0.931 ± 0.016 | 0.767 ± 0.012 | 0.0164 ± 0.0084 | 0.0539 ± 0.0262 | 0.841 ± 0.060 | 0.675 ± 0.037 | 0.915 ± 0.038 | 0.794 ± 0.036 | 0.974 ± 0.015 | 0.908 ± 0.034 |
+| Ji (whole blood) | Zero-shot, sc pseudobulk input (ref.) | 0.932 ± 0.006 | 0.979 ± 0.003 | 0.836 ± 0.011 | 0.0231 ± 0.0015 | 0.0776 ± 0.0041 | 0.893 ± 0.017 | 0.538 ± 0.020 | 0.944 ± 0.009 | 0.645 ± 0.022 | 0.980 ± 0.003 | 0.774 ± 0.021 |
+
+#### real side R2, no depth matching
+
+| cohort | set | AUC kNN | AUC RF | 1-NN acc | MMD² (σ=m) | MMD² (σ sum) |
+|---|---|---:|---:|---:|---:|---:|
+| ImMicro | Real vs real control (R2 vs R3) | 0.502 ± 0.001 | 0.496 ± 0.002 | 0.498 ± 0.001 | -2.6e-05 ± 0.0000 | -6.1e-05 ± 0.0000 |
+| ImMicro | Zero-shot, real bulk | 0.997 ± 0.000 | 0.998 ± 0.000 | 0.981 ± 0.001 | 0.1026 ± 0.0005 | 0.2487 ± 0.0010 |
+| ImMicro | Zero-shot + calibration | 0.973 ± 0.002 | 0.986 ± 0.001 | 0.909 ± 0.003 | 0.0430 ± 0.0024 | 0.1199 ± 0.0067 |
+| ImMicro | scSemiProfiler, 5 donors | 0.999 ± 0.000 | 1.000 ± 0.000 | 0.996 ± 0.000 | 0.1865 ± 0.0092 | 0.4991 ± 0.0241 |
+| ImMicro | LoRA, 5 donors | 0.943 ± 0.003 | 0.960 ± 0.002 | 0.839 ± 0.008 | 0.0231 ± 0.0045 | 0.0677 ± 0.0127 |
+| ImMicro | Zero-shot, sc pseudobulk input (ref.) | 0.967 ± 0.001 | 0.989 ± 0.000 | 0.905 ± 0.001 | 0.0446 ± 0.0003 | 0.1325 ± 0.0008 |
+| Sepsis | Real vs real control (R2 vs R3) | 0.501 ± 0.001 | 0.495 ± 0.004 | 0.503 ± 0.001 | -3.9e-05 ± 0.0000 | -0.0001 ± 0.0000 |
+| Sepsis | Zero-shot, real bulk | 0.997 ± 0.000 | 0.999 ± 0.000 | 0.990 ± 0.000 | 0.1914 ± 0.0035 | 0.4280 ± 0.0064 |
+| Sepsis | Zero-shot + calibration | 0.984 ± 0.001 | 0.988 ± 0.002 | 0.940 ± 0.006 | 0.0404 ± 0.0125 | 0.1242 ± 0.0343 |
+| Sepsis | scSemiProfiler, 5 donors | 0.998 ± 0.001 | 1.000 ± 0.000 | 0.992 ± 0.002 | 0.1666 ± 0.0297 | 0.4743 ± 0.0940 |
+| Sepsis | LoRA, 5 donors | 0.936 ± 0.010 | 0.942 ± 0.007 | 0.840 ± 0.017 | 0.0210 ± 0.0073 | 0.0606 ± 0.0187 |
+| Sepsis | Zero-shot, sc pseudobulk input (ref.) | 0.980 ± 0.001 | 0.984 ± 0.001 | 0.929 ± 0.001 | 0.0285 ± 0.0006 | 0.0886 ± 0.0017 |
+| Imperial | Real vs real control (R2 vs R3) | 0.498 ± 0.003 | 0.501 ± 0.004 | 0.499 ± 0.002 | -1.3e-05 ± 0.0000 | -3.5e-05 ± 0.0000 |
+| Imperial | Zero-shot, real bulk | 0.995 ± 0.000 | 0.998 ± 0.000 | 0.981 ± 0.001 | 0.2425 ± 0.0042 | 0.5198 ± 0.0078 |
+| Imperial | Zero-shot + calibration | 0.961 ± 0.003 | 0.967 ± 0.002 | 0.877 ± 0.005 | 0.0218 ± 0.0021 | 0.0673 ± 0.0062 |
+| Imperial | scSemiProfiler, 5 donors | 0.999 ± 0.000 | 1.000 ± 0.000 | 0.995 ± 0.001 | 0.1372 ± 0.0079 | 0.3741 ± 0.0278 |
+| Imperial | LoRA, 5 donors | 0.918 ± 0.011 | 0.916 ± 0.012 | 0.804 ± 0.016 | 0.0156 ± 0.0052 | 0.0443 ± 0.0147 |
+| Imperial | Zero-shot, sc pseudobulk input (ref.) | 0.959 ± 0.001 | 0.966 ± 0.001 | 0.870 ± 0.001 | 0.0174 ± 0.0002 | 0.0503 ± 0.0005 |
+| Giroux | Real vs real control (R2 vs R3) | 0.495 ± 0.005 | 0.496 ± 0.016 | 0.506 ± 0.011 | -1.9e-05 ± 0.0001 | -3.3e-06 ± 0.0003 |
+| Giroux | Zero-shot, real bulk | 0.993 ± 0.002 | 0.996 ± 0.001 | 0.971 ± 0.004 | 0.1039 ± 0.0089 | 0.2867 ± 0.0200 |
+| Giroux | Zero-shot + calibration | 0.942 ± 0.007 | 0.947 ± 0.005 | 0.861 ± 0.011 | 0.0257 ± 0.0042 | 0.0838 ± 0.0117 |
+| Giroux | scSemiProfiler, 5 donors | 0.997 ± 0.001 | 0.999 ± 0.001 | 0.987 ± 0.002 | 0.0743 ± 0.0084 | 0.2210 ± 0.0255 |
+| Giroux | LoRA, 5 donors | 0.893 ± 0.009 | 0.905 ± 0.010 | 0.793 ± 0.016 | 0.0178 ± 0.0053 | 0.0584 ± 0.0155 |
+| Giroux | Zero-shot, sc pseudobulk input (ref.) | 0.931 ± 0.002 | 0.940 ± 0.004 | 0.839 ± 0.007 | 0.0215 ± 0.0020 | 0.0738 ± 0.0067 |
+| Ji (whole blood) | Real vs real control (R2 vs R3) | 0.491 ± 0.006 | 0.489 ± 0.011 | 0.495 ± 0.005 | -2.8e-06 ± 0.0001 | -2.6e-05 ± 0.0002 |
+| Ji (whole blood) | Zero-shot, real bulk | 0.999 ± 0.001 | 1.000 ± 0.000 | 0.996 ± 0.001 | 0.2734 ± 0.0038 | 0.6915 ± 0.0081 |
+| Ji (whole blood) | Zero-shot + calibration | 0.998 ± 0.000 | 0.999 ± 0.000 | 0.992 ± 0.001 | 0.1382 ± 0.0083 | 0.4002 ± 0.0218 |
+| Ji (whole blood) | scSemiProfiler, 5 donors | 0.996 ± 0.001 | 0.999 ± 0.000 | 0.991 ± 0.003 | 0.1527 ± 0.0078 | 0.4520 ± 0.0267 |
+| Ji (whole blood) | LoRA, 5 donors | 0.955 ± 0.015 | 0.956 ± 0.011 | 0.874 ± 0.026 | 0.0408 ± 0.0211 | 0.1276 ± 0.0645 |
+| Ji (whole blood) | Zero-shot, sc pseudobulk input (ref.) | 0.962 ± 0.003 | 0.980 ± 0.003 | 0.896 ± 0.006 | 0.0256 ± 0.0009 | 0.0848 ± 0.0029 |
+
+#### real side R2, depth matched
+
+| cohort | set | AUC kNN | AUC RF | 1-NN acc | MMD² (σ=m) | MMD² (σ sum) |
+|---|---|---:|---:|---:|---:|---:|
+| ImMicro | Real vs real control (R2 vs R3) | 0.502 ± 0.001 | 0.497 ± 0.002 | 0.497 ± 0.001 | -2.9e-05 ± 0.0000 | -6.9e-05 ± 0.0000 |
+| ImMicro | Zero-shot, real bulk | 0.996 ± 0.000 | 0.998 ± 0.000 | 0.978 ± 0.001 | 0.0964 ± 0.0005 | 0.2309 ± 0.0012 |
+| ImMicro | Zero-shot + calibration | 0.940 ± 0.003 | 0.976 ± 0.001 | 0.840 ± 0.004 | 0.0271 ± 0.0012 | 0.0795 ± 0.0033 |
+| ImMicro | scSemiProfiler, 5 donors | 0.988 ± 0.001 | 0.999 ± 0.000 | 0.964 ± 0.004 | 0.0910 ± 0.0046 | 0.2564 ± 0.0117 |
+| ImMicro | LoRA, 5 donors | 0.877 ± 0.007 | 0.938 ± 0.004 | 0.735 ± 0.009 | 0.0117 ± 0.0018 | 0.0369 ± 0.0052 |
+| ImMicro | Zero-shot, sc pseudobulk input (ref.) | 0.936 ± 0.001 | 0.982 ± 0.001 | 0.843 ± 0.001 | 0.0301 ± 0.0002 | 0.0950 ± 0.0004 |
+| Sepsis | Real vs real control (R2 vs R3) | 0.499 ± 0.001 | 0.495 ± 0.007 | 0.502 ± 0.003 | -3.9e-05 ± 0.0000 | -0.0001 ± 0.0000 |
+| Sepsis | Zero-shot, real bulk | 0.997 ± 0.000 | 0.999 ± 0.000 | 0.989 ± 0.000 | 0.1931 ± 0.0031 | 0.4345 ± 0.0055 |
+| Sepsis | Zero-shot + calibration | 0.961 ± 0.004 | 0.974 ± 0.004 | 0.887 ± 0.009 | 0.0318 ± 0.0119 | 0.0952 ± 0.0330 |
+| Sepsis | scSemiProfiler, 5 donors | 0.996 ± 0.001 | 0.999 ± 0.000 | 0.984 ± 0.003 | 0.1061 ± 0.0254 | 0.3122 ± 0.0766 |
+| Sepsis | LoRA, 5 donors | 0.888 ± 0.014 | 0.914 ± 0.011 | 0.784 ± 0.015 | 0.0161 ± 0.0078 | 0.0459 ± 0.0199 |
+| Sepsis | Zero-shot, sc pseudobulk input (ref.) | 0.948 ± 0.001 | 0.966 ± 0.000 | 0.860 ± 0.003 | 0.0193 ± 0.0003 | 0.0589 ± 0.0008 |
+| Imperial | Real vs real control (R2 vs R3) | 0.498 ± 0.002 | 0.499 ± 0.004 | 0.499 ± 0.001 | -1.4e-05 ± 0.0000 | -3.6e-05 ± 0.0000 |
+| Imperial | Zero-shot, real bulk | 0.994 ± 0.000 | 0.997 ± 0.000 | 0.974 ± 0.001 | 0.2235 ± 0.0044 | 0.4752 ± 0.0085 |
+| Imperial | Zero-shot + calibration | 0.952 ± 0.003 | 0.967 ± 0.002 | 0.865 ± 0.003 | 0.0238 ± 0.0020 | 0.0721 ± 0.0061 |
+| Imperial | scSemiProfiler, 5 donors | 0.997 ± 0.000 | 0.999 ± 0.000 | 0.987 ± 0.002 | 0.1172 ± 0.0126 | 0.3281 ± 0.0362 |
+| Imperial | LoRA, 5 donors | 0.882 ± 0.011 | 0.903 ± 0.008 | 0.764 ± 0.009 | 0.0101 ± 0.0022 | 0.0292 ± 0.0067 |
+| Imperial | Zero-shot, sc pseudobulk input (ref.) | 0.949 ± 0.001 | 0.967 ± 0.001 | 0.854 ± 0.002 | 0.0189 ± 0.0003 | 0.0552 ± 0.0007 |
+| Giroux | Real vs real control (R2 vs R3) | 0.493 ± 0.006 | 0.497 ± 0.010 | 0.502 ± 0.009 | -3.7e-05 ± 0.0001 | -7.5e-05 ± 0.0003 |
+| Giroux | Zero-shot, real bulk | 0.989 ± 0.002 | 0.993 ± 0.001 | 0.959 ± 0.007 | 0.0725 ± 0.0131 | 0.2037 ± 0.0336 |
+| Giroux | Zero-shot + calibration | 0.923 ± 0.013 | 0.935 ± 0.012 | 0.827 ± 0.015 | 0.0238 ± 0.0026 | 0.0774 ± 0.0080 |
+| Giroux | scSemiProfiler, 5 donors | 0.997 ± 0.001 | 0.999 ± 0.000 | 0.984 ± 0.002 | 0.0951 ± 0.0090 | 0.3019 ± 0.0270 |
+| Giroux | LoRA, 5 donors | 0.820 ± 0.006 | 0.864 ± 0.010 | 0.715 ± 0.010 | 0.0076 ± 0.0014 | 0.0248 ± 0.0042 |
+| Giroux | Zero-shot, sc pseudobulk input (ref.) | 0.868 ± 0.005 | 0.894 ± 0.014 | 0.762 ± 0.009 | 0.0193 ± 0.0028 | 0.0630 ± 0.0091 |
+| Ji (whole blood) | Real vs real control (R2 vs R3) | 0.493 ± 0.002 | 0.494 ± 0.008 | 0.495 ± 0.008 | -2.3e-05 ± 0.0001 | -9.1e-05 ± 0.0002 |
+| Ji (whole blood) | Zero-shot, real bulk | 0.996 ± 0.000 | 0.997 ± 0.001 | 0.978 ± 0.000 | 0.1397 ± 0.0017 | 0.3838 ± 0.0045 |
+| Ji (whole blood) | Zero-shot + calibration | 0.993 ± 0.001 | 0.994 ± 0.002 | 0.960 ± 0.004 | 0.0421 ± 0.0043 | 0.1409 ± 0.0120 |
+| Ji (whole blood) | scSemiProfiler, 5 donors | 0.996 ± 0.001 | 0.999 ± 0.000 | 0.988 ± 0.003 | 0.1172 ± 0.0087 | 0.3624 ± 0.0258 |
+| Ji (whole blood) | LoRA, 5 donors | 0.898 ± 0.016 | 0.920 ± 0.017 | 0.792 ± 0.016 | 0.0162 ± 0.0084 | 0.0519 ± 0.0254 |
+| Ji (whole blood) | Zero-shot, sc pseudobulk input (ref.) | 0.954 ± 0.003 | 0.977 ± 0.004 | 0.879 ± 0.005 | 0.0224 ± 0.0014 | 0.0734 ± 0.0041 |
+
+## (b) 5-fold pooled design: one scoring call over all cross-fitted pairs
+
+Units: 5. k = 10 reproduction: max |diff| vs score_ft.py = 1.1e-16 over 30 unit × set rows (pooled and per donor, raw and depth matched); scored rows and P/C subsample identical in every arm: True.
+
+| cohort | arm | AUC kNN | AUC RF | 1-NN acc | MMD² (σ=m) | MMD² (σ sum) | E_N (all cells) | P k3 | C k3 | P k5 | C k5 | P k10 | C k10 |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| ImMicro | Real vs real control | 0.505 | 0.507 | 0.497 | -2.4e-05 | -5.6e-05 | 0.383 | 0.833 | 0.931 | 0.905 | 0.987 | 0.963 | 1.000 |
+| ImMicro | Zero-shot, real bulk | 0.996 | 0.998 | 0.982 | 0.1019 | 0.2480 | 4.832 | 0.400 | 0.092 | 0.518 | 0.128 | 0.680 | 0.192 |
+| ImMicro | Zero-shot + calibration | 0.973 | 0.986 | 0.914 | 0.0425 | 0.1193 | 2.263 | 0.713 | 0.343 | 0.807 | 0.419 | 0.905 | 0.527 |
+| ImMicro | sc-only fine-tune (+ calibration) | 0.952 | 0.954 | 0.861 | 0.0342 | 0.0918 | 1.767 | 0.531 | 0.370 | 0.656 | 0.480 | 0.805 | 0.623 |
+| ImMicro | **LoRA, all training pairs** | 0.941 | 0.951 | 0.842 | 0.0264 | 0.0751 | 1.560 | 0.735 | 0.534 | 0.834 | 0.648 | 0.924 | 0.784 |
+| ImMicro | Zero-shot, sc pseudobulk input (ref.) | 0.967 | 0.989 | 0.908 | 0.0433 | 0.1293 | 1.950 | 0.794 | 0.355 | 0.870 | 0.427 | 0.936 | 0.524 |
+| Sepsis | Real vs real control | 0.506 | 0.497 | 0.503 | -3.9e-05 | -0.0001 | 0.110 | 0.868 | 0.922 | 0.936 | 0.987 | 0.980 | 1.000 |
+| Sepsis | Zero-shot, real bulk | 0.998 | 0.999 | 0.991 | 0.1920 | 0.4306 | 5.288 | 0.309 | 0.032 | 0.429 | 0.044 | 0.614 | 0.067 |
+| Sepsis | Zero-shot + calibration | 0.984 | 0.988 | 0.938 | 0.0340 | 0.1064 | 1.994 | 0.611 | 0.226 | 0.754 | 0.305 | 0.903 | 0.431 |
+| Sepsis | sc-only fine-tune (+ calibration) | 0.901 | 0.910 | 0.795 | 0.0123 | 0.0361 | 1.402 | 0.582 | 0.414 | 0.715 | 0.548 | 0.863 | 0.726 |
+| Sepsis | **LoRA, all training pairs** | 0.886 | 0.901 | 0.785 | 0.0089 | 0.0274 | 1.196 | 0.684 | 0.538 | 0.812 | 0.672 | 0.926 | 0.829 |
+| Sepsis | Zero-shot, sc pseudobulk input (ref.) | 0.980 | 0.984 | 0.930 | 0.0285 | 0.0889 | 1.701 | 0.631 | 0.249 | 0.773 | 0.335 | 0.915 | 0.480 |
+| Imperial | Real vs real control | 0.496 | 0.504 | 0.499 | -1.4e-05 | -3.6e-05 | 0.279 | 0.837 | 0.916 | 0.910 | 0.985 | 0.967 | 1.000 |
+| Imperial | Zero-shot, real bulk | 0.995 | 0.998 | 0.981 | 0.2496 | 0.5309 | 4.052 | 0.207 | 0.050 | 0.311 | 0.072 | 0.489 | 0.112 |
+| Imperial | Zero-shot + calibration | 0.961 | 0.967 | 0.881 | 0.0197 | 0.0600 | 1.194 | 0.589 | 0.377 | 0.709 | 0.490 | 0.842 | 0.639 |
+| Imperial | sc-only fine-tune (+ calibration) | 0.902 | 0.919 | 0.783 | 0.0252 | 0.0672 | 1.408 | 0.414 | 0.323 | 0.534 | 0.418 | 0.691 | 0.555 |
+| Imperial | **LoRA, all training pairs** | 0.866 | 0.870 | 0.751 | 0.0112 | 0.0309 | 1.133 | 0.606 | 0.560 | 0.719 | 0.697 | 0.847 | 0.850 |
+| Imperial | Zero-shot, sc pseudobulk input (ref.) | 0.961 | 0.965 | 0.874 | 0.0174 | 0.0501 | 1.102 | 0.585 | 0.416 | 0.711 | 0.541 | 0.843 | 0.704 |
+| Giroux | Real vs real control | 0.501 | 0.490 | 0.501 | -1.6e-05 | -4.5e-05 | 0.277 | 0.978 | 0.987 | 0.992 | 0.998 | 0.998 | 1.000 |
+| Giroux | Zero-shot, real bulk | 0.996 | 0.998 | 0.985 | 0.1231 | 0.3187 | 4.481 | 0.685 | 0.168 | 0.790 | 0.209 | 0.909 | 0.281 |
+| Giroux | Zero-shot + calibration | 0.945 | 0.946 | 0.860 | 0.0243 | 0.0744 | 2.259 | 0.914 | 0.688 | 0.965 | 0.797 | 0.995 | 0.906 |
+| Giroux | sc-only fine-tune (+ calibration) | 0.898 | 0.907 | 0.796 | 0.0240 | 0.0682 | 2.235 | 0.884 | 0.703 | 0.943 | 0.801 | 0.991 | 0.908 |
+| Giroux | **LoRA, all training pairs** | 0.881 | 0.888 | 0.785 | 0.0177 | 0.0531 | 1.699 | 0.950 | 0.769 | 0.982 | 0.851 | 0.998 | 0.928 |
+| Giroux | Zero-shot, sc pseudobulk input (ref.) | 0.931 | 0.937 | 0.838 | 0.0235 | 0.0741 | 1.872 | 0.948 | 0.659 | 0.979 | 0.756 | 0.998 | 0.865 |
+| Ji (whole blood) | Real vs real control | 0.494 | 0.506 | 0.499 | -1.6e-05 | -4.1e-05 | 0.074 | 0.963 | 0.975 | 0.983 | 0.997 | 0.996 | 1.000 |
+| Ji (whole blood) | Zero-shot, real bulk | 0.999 | 1.000 | 0.996 | 0.2746 | 0.6818 | 10.125 | 0.339 | 0.025 | 0.487 | 0.035 | 0.717 | 0.055 |
+| Ji (whole blood) | Zero-shot + calibration | 0.998 | 0.999 | 0.992 | 0.1317 | 0.3753 | 7.076 | 0.371 | 0.050 | 0.527 | 0.071 | 0.740 | 0.106 |
+| Ji (whole blood) | sc-only fine-tune (+ calibration) | 0.997 | 0.998 | 0.983 | 0.1536 | 0.4390 | 7.215 | 0.328 | 0.059 | 0.464 | 0.079 | 0.694 | 0.113 |
+| Ji (whole blood) | **LoRA, all training pairs** | 0.908 | 0.909 | 0.803 | 0.0212 | 0.0637 | 1.948 | 0.844 | 0.629 | 0.919 | 0.734 | 0.971 | 0.857 |
+| Ji (whole blood) | Zero-shot, sc pseudobulk input (ref.) | 0.960 | 0.980 | 0.891 | 0.0255 | 0.0834 | 1.687 | 0.896 | 0.468 | 0.946 | 0.562 | 0.985 | 0.685 |
+
+### (b) 5-fold pooled design: one scoring call over all cross-fitted pairs: is LoRA better than each comparator? (✓ better / ✗ not)
+
+| cohort | LoRA vs | AUC kNN | AUC RF | 1-NN acc | MMD² (σ=m) | MMD² (σ sum) | E_N (all cells) | P k3 | C k3 | P k5 | C k5 | P k10 | C k10 | E_N (P/C cells) |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| ImMicro | Zero-shot, real bulk | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
+| ImMicro | Zero-shot + calibration | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
+| ImMicro | sc-only fine-tune (+ calibration) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
+| Sepsis | Zero-shot, real bulk | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
+| Sepsis | Zero-shot + calibration | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
+| Sepsis | sc-only fine-tune (+ calibration) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
+| Imperial | Zero-shot, real bulk | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
+| Imperial | Zero-shot + calibration | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
+| Imperial | sc-only fine-tune (+ calibration) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
+| Giroux | Zero-shot, real bulk | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
+| Giroux | Zero-shot + calibration | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
+| Giroux | sc-only fine-tune (+ calibration) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
+| Ji (whole blood) | Zero-shot, real bulk | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
+| Ji (whole blood) | Zero-shot + calibration | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
+| Ji (whole blood) | sc-only fine-tune (+ calibration) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
+
+Exceptions (LoRA worse; ties are not counted as worse):
+
+- none
+
+### (b) 5-fold pooled design: one scoring call over all cross-fitted pairs: Kendall τ of each metric's arm ranking with the coverage / precision (k = 10) ranking (arms only; few-pair: means over draws, E_N medians)
+
+| cohort | vs | AUC kNN | AUC RF | 1-NN acc | MMD² (σ=m) | MMD² (σ sum) | E_N (all cells) | P k3 | C k3 | P k5 | C k5 | P k10 | C k10 | E_N (P/C cells) |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| ImMicro | coverage k10 | 0.80 | 1.00 | 0.80 | 1.00 | 1.00 | 0.80 | 0.20 | 0.80 | 0.20 | 0.80 | 0.20 | 1.00 | 0.80 |
+| ImMicro | precision k10 | 0.40 | 0.20 | 0.40 | 0.20 | 0.20 | 0.40 | 1.00 | 0.40 | 1.00 | 0.40 | 1.00 | 0.20 | 0.40 |
+| Sepsis | coverage k10 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 0.60 | 1.00 | 0.60 | 1.00 | 0.60 | 1.00 | 1.00 |
+| Sepsis | precision k10 | 0.60 | 0.60 | 0.60 | 0.60 | 0.60 | 0.60 | 1.00 | 0.60 | 1.00 | 0.60 | 1.00 | 0.60 | 0.60 |
+| Imperial | coverage k10 | 0.60 | 0.60 | 0.60 | 1.00 | 1.00 | 0.80 | 0.80 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 0.80 |
+| Imperial | precision k10 | 0.60 | 0.60 | 0.60 | 1.00 | 1.00 | 0.80 | 0.80 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 0.80 |
+| Giroux | coverage k10 | 0.80 | 0.80 | 0.80 | 0.60 | 0.80 | 0.60 | 0.40 | 1.00 | 0.40 | 1.00 | 0.40 | 1.00 | 0.40 |
+| Giroux | precision k10 | 0.60 | 0.60 | 0.60 | 0.80 | 0.60 | 0.80 | 1.00 | 0.40 | 1.00 | 0.40 | 1.00 | 0.40 | 0.60 |
+| Ji (whole blood) | coverage k10 | 1.00 | 1.00 | 1.00 | 0.80 | 0.80 | 0.60 | 0.40 | 1.00 | 0.40 | 1.00 | 0.40 | 1.00 | 0.80 |
+| Ji (whole blood) | precision k10 | 0.40 | 0.40 | 0.40 | 0.60 | 0.60 | 0.80 | 1.00 | 0.40 | 1.00 | 0.40 | 1.00 | 0.40 | 0.60 |
+
+### (b) 5-fold pooled design: one scoring call over all cross-fitted pairs: every variant (real side R1 / R2, raw / depth matched)
+
+#### real side R1, no depth matching
+
+| cohort | set | AUC kNN | AUC RF | 1-NN acc | MMD² (σ=m) | MMD² (σ sum) | P k3 | C k3 | P k5 | C k5 | P k10 | C k10 | E_N (all cells) | E_N (P/C cells) |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| ImMicro | Real vs real control (R1 vs R2) | 0.504 | 0.581 | 0.499 | 3.1e-05 | 0.0002 | 0.833 | 0.931 | 0.905 | 0.987 | 0.963 | 1.000 | 0.383 | 0.515 |
+| ImMicro | Zero-shot, real bulk | 0.995 | 0.998 | 0.978 | 0.0998 | 0.2400 | 0.400 | 0.092 | 0.518 | 0.128 | 0.680 | 0.192 | 4.832 | 24.929 |
+| ImMicro | Zero-shot + calibration | 0.967 | 0.986 | 0.899 | 0.0423 | 0.1194 | 0.713 | 0.343 | 0.807 | 0.419 | 0.905 | 0.527 | 2.263 | 2.847 |
+| ImMicro | sc-only fine-tune (+ calibration) | 0.950 | 0.953 | 0.861 | 0.0338 | 0.0905 | 0.531 | 0.370 | 0.656 | 0.480 | 0.805 | 0.623 | 1.767 | 1.987 |
+| ImMicro | LoRA, all training pairs | 0.936 | 0.954 | 0.829 | 0.0264 | 0.0759 | 0.735 | 0.534 | 0.834 | 0.648 | 0.924 | 0.784 | 1.560 | 1.777 |
+| ImMicro | Zero-shot, sc pseudobulk input (ref.) | 0.959 | 0.989 | 0.894 | 0.0435 | 0.1311 | 0.794 | 0.355 | 0.870 | 0.427 | 0.936 | 0.524 | 1.950 | 2.086 |
+| Sepsis | Real vs real control (R1 vs R2) | 0.517 | 0.587 | 0.508 | 1.4e-05 | 0.0001 | 0.868 | 0.922 | 0.936 | 0.987 | 0.980 | 1.000 | 0.110 | 0.224 |
+| Sepsis | Zero-shot, real bulk | 0.997 | 0.999 | 0.990 | 0.1897 | 0.4223 | 0.309 | 0.032 | 0.429 | 0.044 | 0.614 | 0.067 | 5.288 | 19.582 |
+| Sepsis | Zero-shot + calibration | 0.981 | 0.987 | 0.933 | 0.0343 | 0.1077 | 0.611 | 0.226 | 0.754 | 0.305 | 0.903 | 0.431 | 1.994 | 1.492 |
+| Sepsis | sc-only fine-tune (+ calibration) | 0.907 | 0.910 | 0.810 | 0.0124 | 0.0366 | 0.582 | 0.414 | 0.715 | 0.548 | 0.863 | 0.726 | 1.402 | 1.151 |
+| Sepsis | LoRA, all training pairs | 0.890 | 0.908 | 0.790 | 0.0094 | 0.0294 | 0.684 | 0.538 | 0.812 | 0.672 | 0.926 | 0.829 | 1.196 | 1.011 |
+| Sepsis | Zero-shot, sc pseudobulk input (ref.) | 0.978 | 0.983 | 0.925 | 0.0290 | 0.0909 | 0.631 | 0.249 | 0.773 | 0.335 | 0.915 | 0.480 | 1.701 | 1.246 |
+| Imperial | Real vs real control (R1 vs R2) | 0.505 | 0.557 | 0.503 | -5.1e-06 | 2.5e-05 | 0.837 | 0.916 | 0.910 | 0.985 | 0.967 | 1.000 | 0.279 | 0.235 |
+| Imperial | Zero-shot, real bulk | 0.995 | 0.998 | 0.979 | 0.2486 | 0.5258 | 0.207 | 0.050 | 0.311 | 0.072 | 0.489 | 0.112 | 4.052 | 6.385 |
+| Imperial | Zero-shot + calibration | 0.959 | 0.966 | 0.874 | 0.0199 | 0.0610 | 0.589 | 0.377 | 0.709 | 0.490 | 0.842 | 0.639 | 1.194 | 1.132 |
+| Imperial | sc-only fine-tune (+ calibration) | 0.903 | 0.916 | 0.788 | 0.0255 | 0.0674 | 0.414 | 0.323 | 0.534 | 0.418 | 0.691 | 0.555 | 1.408 | 1.393 |
+| Imperial | LoRA, all training pairs | 0.864 | 0.871 | 0.751 | 0.0113 | 0.0313 | 0.606 | 0.560 | 0.719 | 0.697 | 0.847 | 0.850 | 1.133 | 1.080 |
+| Imperial | Zero-shot, sc pseudobulk input (ref.) | 0.957 | 0.965 | 0.867 | 0.0175 | 0.0508 | 0.585 | 0.416 | 0.711 | 0.541 | 0.843 | 0.704 | 1.102 | 1.022 |
+| Giroux | Real vs real control (R1 vs R2) | 0.566 | 0.844 | 0.527 | 0.0006 | 0.0041 | 0.978 | 0.987 | 0.992 | 0.998 | 0.998 | 1.000 | 0.277 | 0.393 |
+| Giroux | Zero-shot, real bulk | 0.984 | 0.997 | 0.954 | 0.1198 | 0.3019 | 0.685 | 0.168 | 0.790 | 0.209 | 0.909 | 0.281 | 4.481 | 6.922 |
+| Giroux | Zero-shot + calibration | 0.916 | 0.962 | 0.786 | 0.0242 | 0.0757 | 0.914 | 0.688 | 0.965 | 0.797 | 0.995 | 0.906 | 2.259 | 2.152 |
+| Giroux | sc-only fine-tune (+ calibration) | 0.882 | 0.931 | 0.753 | 0.0252 | 0.0724 | 0.884 | 0.703 | 0.943 | 0.801 | 0.991 | 0.908 | 2.235 | 2.057 |
+| Giroux | LoRA, all training pairs | 0.840 | 0.929 | 0.718 | 0.0198 | 0.0652 | 0.950 | 0.769 | 0.982 | 0.851 | 0.998 | 0.928 | 1.699 | 1.737 |
+| Giroux | Zero-shot, sc pseudobulk input (ref.) | 0.892 | 0.955 | 0.776 | 0.0241 | 0.0804 | 0.948 | 0.659 | 0.979 | 0.756 | 0.998 | 0.865 | 1.872 | 1.557 |
+| Ji (whole blood) | Real vs real control (R1 vs R2) | 0.541 | 0.765 | 0.523 | 0.0001 | 0.0010 | 0.963 | 0.975 | 0.983 | 0.997 | 0.996 | 1.000 | 0.074 | 0.209 |
+| Ji (whole blood) | Zero-shot, real bulk | 0.998 | 0.999 | 0.993 | 0.2725 | 0.6668 | 0.339 | 0.025 | 0.487 | 0.035 | 0.717 | 0.055 | 10.125 | — †1 |
+| Ji (whole blood) | Zero-shot + calibration | 0.996 | 0.999 | 0.988 | 0.1313 | 0.3697 | 0.371 | 0.050 | 0.527 | 0.071 | 0.740 | 0.106 | 7.076 | 9.766 |
+| Ji (whole blood) | sc-only fine-tune (+ calibration) | 0.995 | 0.998 | 0.981 | 0.1525 | 0.4312 | 0.328 | 0.059 | 0.464 | 0.079 | 0.694 | 0.113 | 7.215 | 5.990 |
+| Ji (whole blood) | LoRA, all training pairs | 0.902 | 0.937 | 0.790 | 0.0220 | 0.0679 | 0.844 | 0.629 | 0.919 | 0.734 | 0.971 | 0.857 | 1.948 | 1.439 |
+| Ji (whole blood) | Zero-shot, sc pseudobulk input (ref.) | 0.940 | 0.983 | 0.855 | 0.0264 | 0.0883 | 0.896 | 0.468 | 0.946 | 0.562 | 0.985 | 0.685 | 1.687 | 1.045 |
+
+#### real side R1, depth matched
+
+| cohort | set | AUC kNN | AUC RF | 1-NN acc | MMD² (σ=m) | MMD² (σ sum) | P k3 | C k3 | P k5 | C k5 | P k10 | C k10 |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| ImMicro | Real vs real control (R1 vs R2) | 0.507 | 0.577 | 0.498 | 2.9e-05 | 0.0002 | 0.829 | 0.928 | 0.905 | 0.987 | 0.963 | 1.000 |
+| ImMicro | Zero-shot, real bulk | 0.995 | 0.998 | 0.975 | 0.0944 | 0.2250 | 0.395 | 0.100 | 0.503 | 0.140 | 0.660 | 0.218 |
+| ImMicro | Zero-shot + calibration | 0.934 | 0.976 | 0.829 | 0.0264 | 0.0792 | 0.807 | 0.550 | 0.880 | 0.652 | 0.944 | 0.769 |
+| ImMicro | sc-only fine-tune (+ calibration) | 0.870 | 0.917 | 0.752 | 0.0152 | 0.0422 | 0.739 | 0.696 | 0.826 | 0.809 | 0.914 | 0.919 |
+| ImMicro | LoRA, all training pairs | 0.861 | 0.931 | 0.720 | 0.0115 | 0.0366 | 0.848 | 0.811 | 0.909 | 0.900 | 0.961 | 0.962 |
+| ImMicro | Zero-shot, sc pseudobulk input (ref.) | 0.927 | 0.983 | 0.834 | 0.0296 | 0.0952 | 0.867 | 0.526 | 0.919 | 0.627 | 0.965 | 0.741 |
+| Sepsis | Real vs real control (R1 vs R2) | 0.514 | 0.581 | 0.503 | 1.3e-05 | 9.7e-05 | 0.868 | 0.921 | 0.933 | 0.985 | 0.980 | 1.000 |
+| Sepsis | Zero-shot, real bulk | 0.997 | 0.999 | 0.990 | 0.1916 | 0.4298 | 0.278 | 0.029 | 0.413 | 0.042 | 0.592 | 0.065 |
+| Sepsis | Zero-shot + calibration | 0.956 | 0.973 | 0.873 | 0.0250 | 0.0761 | 0.664 | 0.413 | 0.790 | 0.532 | 0.911 | 0.683 |
+| Sepsis | sc-only fine-tune (+ calibration) | 0.822 | 0.870 | 0.728 | 0.0058 | 0.0170 | 0.624 | 0.609 | 0.738 | 0.750 | 0.863 | 0.892 |
+| Sepsis | LoRA, all training pairs | 0.806 | 0.861 | 0.711 | 0.0050 | 0.0154 | 0.743 | 0.723 | 0.843 | 0.846 | 0.933 | 0.949 |
+| Sepsis | Zero-shot, sc pseudobulk input (ref.) | 0.943 | 0.963 | 0.853 | 0.0192 | 0.0589 | 0.726 | 0.465 | 0.836 | 0.589 | 0.932 | 0.746 |
+| Imperial | Real vs real control (R1 vs R2) | 0.503 | 0.550 | 0.506 | -7.3e-06 | 1.5e-05 | 0.836 | 0.916 | 0.910 | 0.984 | 0.967 | 1.000 |
+| Imperial | Zero-shot, real bulk | 0.994 | 0.996 | 0.972 | 0.2294 | 0.4815 | 0.239 | 0.068 | 0.365 | 0.095 | 0.566 | 0.147 |
+| Imperial | Zero-shot + calibration | 0.948 | 0.965 | 0.859 | 0.0216 | 0.0647 | 0.645 | 0.418 | 0.758 | 0.538 | 0.864 | 0.688 |
+| Imperial | sc-only fine-tune (+ calibration) | 0.834 | 0.878 | 0.722 | 0.0113 | 0.0315 | 0.480 | 0.451 | 0.605 | 0.593 | 0.756 | 0.771 |
+| Imperial | LoRA, all training pairs | 0.818 | 0.853 | 0.712 | 0.0067 | 0.0188 | 0.671 | 0.675 | 0.779 | 0.810 | 0.888 | 0.937 |
+| Imperial | Zero-shot, sc pseudobulk input (ref.) | 0.947 | 0.967 | 0.853 | 0.0191 | 0.0556 | 0.661 | 0.456 | 0.772 | 0.576 | 0.885 | 0.732 |
+| Giroux | Real vs real control (R1 vs R2) | 0.571 | 0.840 | 0.527 | 0.0006 | 0.0039 | 0.978 | 0.985 | 0.991 | 0.998 | 0.998 | 1.000 |
+| Giroux | Zero-shot, real bulk | 0.965 | 0.997 | 0.923 | 0.0842 | 0.2201 | 0.856 | 0.238 | 0.912 | 0.283 | 0.975 | 0.355 |
+| Giroux | Zero-shot + calibration | 0.904 | 0.946 | 0.778 | 0.0225 | 0.0684 | 0.851 | 0.686 | 0.916 | 0.796 | 0.981 | 0.902 |
+| Giroux | sc-only fine-tune (+ calibration) | 0.887 | 0.921 | 0.772 | 0.0212 | 0.0597 | 0.766 | 0.607 | 0.865 | 0.731 | 0.958 | 0.846 |
+| Giroux | LoRA, all training pairs | 0.782 | 0.863 | 0.684 | 0.0064 | 0.0210 | 0.878 | 0.824 | 0.944 | 0.916 | 0.984 | 0.974 |
+| Giroux | Zero-shot, sc pseudobulk input (ref.) | 0.849 | 0.904 | 0.735 | 0.0209 | 0.0630 | 0.871 | 0.745 | 0.946 | 0.844 | 0.987 | 0.931 |
+| Ji (whole blood) | Real vs real control (R1 vs R2) | 0.540 | 0.756 | 0.520 | 8.7e-05 | 0.0009 | 0.958 | 0.977 | 0.980 | 0.998 | 0.997 | 1.000 |
+| Ji (whole blood) | Zero-shot, real bulk | 0.996 | 0.997 | 0.979 | 0.1435 | 0.3820 | 0.224 | 0.069 | 0.321 | 0.104 | 0.489 | 0.167 |
+| Ji (whole blood) | Zero-shot + calibration | 0.991 | 0.993 | 0.961 | 0.0415 | 0.1353 | 0.344 | 0.162 | 0.481 | 0.227 | 0.670 | 0.347 |
+| Ji (whole blood) | sc-only fine-tune (+ calibration) | 0.975 | 0.978 | 0.914 | 0.0443 | 0.1433 | 0.358 | 0.222 | 0.470 | 0.313 | 0.640 | 0.455 |
+| Ji (whole blood) | LoRA, all training pairs | 0.836 | 0.890 | 0.720 | 0.0069 | 0.0226 | 0.840 | 0.774 | 0.916 | 0.880 | 0.973 | 0.965 |
+| Ji (whole blood) | Zero-shot, sc pseudobulk input (ref.) | 0.937 | 0.979 | 0.844 | 0.0229 | 0.0756 | 0.881 | 0.501 | 0.936 | 0.604 | 0.978 | 0.738 |
+
+#### real side R2, no depth matching
+
+| cohort | set | AUC kNN | AUC RF | 1-NN acc | MMD² (σ=m) | MMD² (σ sum) |
+|---|---|---:|---:|---:|---:|---:|
+| ImMicro | Real vs real control (R2 vs R3) | 0.505 | 0.507 | 0.497 | -2.4e-05 | -5.6e-05 |
+| ImMicro | Zero-shot, real bulk | 0.996 | 0.998 | 0.982 | 0.1019 | 0.2480 |
+| ImMicro | Zero-shot + calibration | 0.973 | 0.986 | 0.914 | 0.0425 | 0.1193 |
+| ImMicro | sc-only fine-tune (+ calibration) | 0.952 | 0.954 | 0.861 | 0.0342 | 0.0918 |
+| ImMicro | LoRA, all training pairs | 0.941 | 0.951 | 0.842 | 0.0264 | 0.0751 |
+| ImMicro | Zero-shot, sc pseudobulk input (ref.) | 0.967 | 0.989 | 0.908 | 0.0433 | 0.1293 |
+| Sepsis | Real vs real control (R2 vs R3) | 0.506 | 0.497 | 0.503 | -3.9e-05 | -0.0001 |
+| Sepsis | Zero-shot, real bulk | 0.998 | 0.999 | 0.991 | 0.1920 | 0.4306 |
+| Sepsis | Zero-shot + calibration | 0.984 | 0.988 | 0.938 | 0.0340 | 0.1064 |
+| Sepsis | sc-only fine-tune (+ calibration) | 0.901 | 0.910 | 0.795 | 0.0123 | 0.0361 |
+| Sepsis | LoRA, all training pairs | 0.886 | 0.901 | 0.785 | 0.0089 | 0.0274 |
+| Sepsis | Zero-shot, sc pseudobulk input (ref.) | 0.980 | 0.984 | 0.930 | 0.0285 | 0.0889 |
+| Imperial | Real vs real control (R2 vs R3) | 0.496 | 0.504 | 0.499 | -1.4e-05 | -3.6e-05 |
+| Imperial | Zero-shot, real bulk | 0.995 | 0.998 | 0.981 | 0.2496 | 0.5309 |
+| Imperial | Zero-shot + calibration | 0.961 | 0.967 | 0.881 | 0.0197 | 0.0600 |
+| Imperial | sc-only fine-tune (+ calibration) | 0.902 | 0.919 | 0.783 | 0.0252 | 0.0672 |
+| Imperial | LoRA, all training pairs | 0.866 | 0.870 | 0.751 | 0.0112 | 0.0309 |
+| Imperial | Zero-shot, sc pseudobulk input (ref.) | 0.961 | 0.965 | 0.874 | 0.0174 | 0.0501 |
+| Giroux | Real vs real control (R2 vs R3) | 0.501 | 0.490 | 0.501 | -1.6e-05 | -4.5e-05 |
+| Giroux | Zero-shot, real bulk | 0.996 | 0.998 | 0.985 | 0.1231 | 0.3187 |
+| Giroux | Zero-shot + calibration | 0.945 | 0.946 | 0.860 | 0.0243 | 0.0744 |
+| Giroux | sc-only fine-tune (+ calibration) | 0.898 | 0.907 | 0.796 | 0.0240 | 0.0682 |
+| Giroux | LoRA, all training pairs | 0.881 | 0.888 | 0.785 | 0.0177 | 0.0531 |
+| Giroux | Zero-shot, sc pseudobulk input (ref.) | 0.931 | 0.937 | 0.838 | 0.0235 | 0.0741 |
+| Ji (whole blood) | Real vs real control (R2 vs R3) | 0.494 | 0.506 | 0.499 | -1.6e-05 | -4.1e-05 |
+| Ji (whole blood) | Zero-shot, real bulk | 0.999 | 1.000 | 0.996 | 0.2746 | 0.6818 |
+| Ji (whole blood) | Zero-shot + calibration | 0.998 | 0.999 | 0.992 | 0.1317 | 0.3753 |
+| Ji (whole blood) | sc-only fine-tune (+ calibration) | 0.997 | 0.998 | 0.983 | 0.1536 | 0.4390 |
+| Ji (whole blood) | LoRA, all training pairs | 0.908 | 0.909 | 0.803 | 0.0212 | 0.0637 |
+| Ji (whole blood) | Zero-shot, sc pseudobulk input (ref.) | 0.960 | 0.980 | 0.891 | 0.0255 | 0.0834 |
+
+#### real side R2, depth matched
+
+| cohort | set | AUC kNN | AUC RF | 1-NN acc | MMD² (σ=m) | MMD² (σ sum) |
+|---|---|---:|---:|---:|---:|---:|
+| ImMicro | Real vs real control (R2 vs R3) | 0.503 | 0.499 | 0.497 | -2.7e-05 | -6.5e-05 |
+| ImMicro | Zero-shot, real bulk | 0.996 | 0.998 | 0.979 | 0.0959 | 0.2308 |
+| ImMicro | Zero-shot + calibration | 0.943 | 0.976 | 0.846 | 0.0265 | 0.0783 |
+| ImMicro | sc-only fine-tune (+ calibration) | 0.876 | 0.917 | 0.760 | 0.0153 | 0.0424 |
+| ImMicro | LoRA, all training pairs | 0.868 | 0.927 | 0.732 | 0.0114 | 0.0353 |
+| ImMicro | Zero-shot, sc pseudobulk input (ref.) | 0.938 | 0.982 | 0.848 | 0.0293 | 0.0930 |
+| Sepsis | Real vs real control (R2 vs R3) | 0.503 | 0.502 | 0.502 | -3.9e-05 | -0.0001 |
+| Sepsis | Zero-shot, real bulk | 0.998 | 0.999 | 0.990 | 0.1937 | 0.4374 |
+| Sepsis | Zero-shot + calibration | 0.959 | 0.974 | 0.880 | 0.0251 | 0.0764 |
+| Sepsis | sc-only fine-tune (+ calibration) | 0.823 | 0.874 | 0.726 | 0.0059 | 0.0173 |
+| Sepsis | LoRA, all training pairs | 0.805 | 0.858 | 0.710 | 0.0048 | 0.0145 |
+| Sepsis | Zero-shot, sc pseudobulk input (ref.) | 0.946 | 0.966 | 0.864 | 0.0192 | 0.0587 |
+| Imperial | Real vs real control (R2 vs R3) | 0.495 | 0.499 | 0.499 | -1.4e-05 | -3.7e-05 |
+| Imperial | Zero-shot, real bulk | 0.995 | 0.997 | 0.974 | 0.2299 | 0.4849 |
+| Imperial | Zero-shot + calibration | 0.950 | 0.966 | 0.864 | 0.0215 | 0.0644 |
+| Imperial | sc-only fine-tune (+ calibration) | 0.837 | 0.882 | 0.724 | 0.0113 | 0.0316 |
+| Imperial | LoRA, all training pairs | 0.817 | 0.853 | 0.709 | 0.0068 | 0.0190 |
+| Imperial | Zero-shot, sc pseudobulk input (ref.) | 0.949 | 0.967 | 0.858 | 0.0192 | 0.0554 |
+| Giroux | Real vs real control (R2 vs R3) | 0.502 | 0.495 | 0.503 | -2.7e-05 | -7.9e-05 |
+| Giroux | Zero-shot, real bulk | 0.992 | 0.995 | 0.968 | 0.0846 | 0.2233 |
+| Giroux | Zero-shot + calibration | 0.927 | 0.935 | 0.840 | 0.0234 | 0.0715 |
+| Giroux | sc-only fine-tune (+ calibration) | 0.894 | 0.917 | 0.788 | 0.0213 | 0.0615 |
+| Giroux | LoRA, all training pairs | 0.800 | 0.844 | 0.716 | 0.0063 | 0.0191 |
+| Giroux | Zero-shot, sc pseudobulk input (ref.) | 0.864 | 0.889 | 0.767 | 0.0219 | 0.0660 |
+| Ji (whole blood) | Real vs real control (R2 vs R3) | 0.494 | 0.501 | 0.493 | -3.0e-05 | -9.7e-05 |
+| Ji (whole blood) | Zero-shot, real bulk | 0.996 | 0.997 | 0.981 | 0.1431 | 0.3844 |
+| Ji (whole blood) | Zero-shot + calibration | 0.992 | 0.993 | 0.962 | 0.0411 | 0.1347 |
+| Ji (whole blood) | sc-only fine-tune (+ calibration) | 0.973 | 0.978 | 0.909 | 0.0446 | 0.1446 |
+| Ji (whole blood) | LoRA, all training pairs | 0.833 | 0.874 | 0.727 | 0.0067 | 0.0206 |
+| Ji (whole blood) | Zero-shot, sc pseudobulk input (ref.) | 0.951 | 0.977 | 0.877 | 0.0227 | 0.0734 |
+
+## Run time per unit (seconds; server_1 CPU, few-pair 6 threads, pooled 8 threads)
+
+| design | cohort | units | samples per unit | mean s | min s | max s |
+|---|---|---:|---:|---:|---:|---:|
+| fewshot | Giroux | 5 | 11 | 245 | 222 | 260 |
+| fewshot | ImMicro | 5 | 102 | 1446 | 1421 | 1479 |
+| fewshot | Imperial | 5 | 143 | 2249 | 2025 | 2350 |
+| fewshot | Ji (whole blood) | 5 | 25 | 330 | 317 | 341 |
+| fewshot | Sepsis | 5 | 77 | 1331 | 1302 | 1364 |
+| pooled | Giroux | 1 | 24 | 461 | 461 | 461 |
+| pooled | ImMicro | 1 | 107 | 1591 | 1591 | 1591 |
+| pooled | Imperial | 1 | 157 | 2449 | 2449 | 2449 |
+| pooled | Ji (whole blood) | 1 | 33 | 387 | 387 | 387 |
+| pooled | Sepsis | 1 | 82 | 1373 | 1373 | 1373 |
+

paper-releases/pbisc-diffusion/updates/20261001T200638Z-extra-realism-metrics/diffusion/dataset/server_1/model8/plans/paired_finetune_intradataset_20260927/scores_extra/_code/README.md

@@ -0,0 +1,19 @@
+# Code versions of tools/extra_metrics.py used for the units in scores_extra/
+
+Each job of `tools/run_extra_unit.sh` starts a new Python process, which loads `tools/extra_metrics.py` at that moment.
+The file was edited three times while the two streams ran (2026-10-02 00:01–05:00 KST). The metric code is identical
+in all three versions; only the table collector and a per-unit lock changed:
+
+| version | sha256 (first 12) | change | units |
+|---|---|---|---|
+| v1 | ca3804cffbaa | as launched | few-pair Giroux, few-pair sepsis |
+| v2 | b65e022452d8 | collector only (E_N columns dropped from depth-matched tables; Dif / LB / UB added to the wide table) | few-pair Ji |
+| v3 | 4e69c029476c | + `.in_progress` lock so two streams never compute the same unit | every other unit |
+
+`code_versions.csv` maps each unit to the version it ran. The `code_sha256` block of each `manifest.json` was hashed when
+the unit **finished**, so for long units it can show a later file; `extra_metrics_py_sha256_ran` in the same manifest is
+the version the process loaded. v1 was reconstructed by reverting the v2 edit of the saved v2 file; its sha256 equals the
+value recorded by the v1 units' manifests (ca3804cffbaa…).
+
+The current `tools/extra_metrics.py` (v4) differs from v3 only in bookkeeping: `--collect` delegates to
+`tools/collect_extra_metrics.py` (the collector code moved there), and the code sha256 is taken at process start.

paper-releases/pbisc-diffusion/updates/20261001T200638Z-extra-realism-metrics/diffusion/dataset/server_1/model8/plans/paired_finetune_intradataset_20260927/tools/run_extra_unit.sh

@@ -0,0 +1,34 @@
+#!/usr/bin/env bash
+# Run a list of tools/extra_metrics.py jobs (CPU only) in one transient systemd user unit on server_1.
+#   MemoryMax MEM, MemorySwapMax 0, CPUQuota THREADS x 100 %, Nice 10, CUDA hidden, no bytecode written.
+#   Refuses to start when MemAvailable < 60 GiB. Does not wait; the unit's log is logs/NAME_<stamp>.log.
+#   Jobs run one after another; a failed job is logged ("JOB_EXIT <rc> <args>") and the next job starts.
+# usage: run_extra_unit.sh NAME THREADS MEM JOBFILE FETCH_DIR
+#   JOBFILE: one extra_metrics.py argument line per row (--threads T in a line overrides THREADS for that job), e.g. "--design fewshot --dataset ji2025_chili --draws 0,1,2,3,4"
+set -uo pipefail
+export XDG_RUNTIME_DIR="${XDG_RUNTIME_DIR:-/run/user/$(id -u)}"
+export DBUS_SESSION_BUS_ADDRESS="${DBUS_SESSION_BUS_ADDRESS:-unix:path=$XDG_RUNTIME_DIR/bus}"
+TASK=/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927
+PY=/home/sj_server_1/miniforge3/envs/cxg_corpus/bin/python
+name=$1; thr=$2; mem=$3; jobs=$4; fetch=$5
+(( thr >= 1 && thr <= 8 )) || { echo "THREADS must be 1..8" >&2; exit 2; }
+[[ -s $jobs ]] || { echo "no job file $jobs" >&2; exit 2; }
+avail=$(awk '/MemAvailable/ {print int($2/1048576)}' /proc/meminfo)
+(( avail >= 60 )) || { echo "MemAvailable ${avail} GiB < 60; not starting $name" >&2; exit 3; }
+mkdir -p "$TASK/logs" "$fetch"
+stamp=$(date +%Y%m%dT%H%M%S)
+unit=extra-${name}-${stamp}
+log=$TASK/logs/${name}_${stamp}.log
+cp "$jobs" "$TASK/logs/${name}_${stamp}.jobs"
+echo -e "$(date -Is)\tstart\t$unit\t$(hostname -s)\tMemAvailable=${avail}GiB\tthreads=$thr\tmem=$mem\t$jobs" >> "$TASK/logs/units.tsv"
+systemd-run --user --unit="$unit" --working-directory="$TASK" \
+  -p MemoryMax="$mem" -p MemorySwapMax=0 -p CPUQuota=$((thr * 100))% -p Nice=10 \
+  -p StandardOutput=append:"$log" -p StandardError=append:"$log" \
+  --setenv=CUDA_VISIBLE_DEVICES= --setenv=PYTHONNOUSERSITE=1 --setenv=PYTHONDONTWRITEBYTECODE=1 \
+  --setenv=PYTHONUNBUFFERED=1 --setenv=OMP_NUM_THREADS="$thr" --setenv=OPENBLAS_NUM_THREADS="$thr" \
+  --setenv=MKL_NUM_THREADS="$thr" --setenv=NUMBA_NUM_THREADS="$thr" \
+  /bin/bash -c 'while IFS= read -r line; do [[ -z $line || $line == \#* ]] && continue; t0=$(date +%s);
+      echo "JOB_START $(date -Is) $line"; [[ $line == *--threads* ]] || line="$line --threads '"$thr"'";
+      '"$PY"' tools/extra_metrics.py $line --fetch-dir '"$fetch"';
+      echo "JOB_EXIT $? $(( $(date +%s) - t0 ))s $line"; done < '"$TASK/logs/${name}_${stamp}.jobs"'; echo UNIT_EXIT 0' >/dev/null
+echo "unit=$unit log=$log"