Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261005T142225Z-rare-states-scratch-vs-pretrained/diffusion/dataset/server_1/model8/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/REPORT.md

@@ -0,0 +1,407 @@
+# Rare-state generation: from scratch vs pretrained (LoRA / full FT)
+
+PBISC-Diffusion, Model 8. Ye 2026 paired sepsis cohort (paediatric + adult PBMC bulk/sc). 2026-10-05.
+
+Sources:
+- Screen outputs in `out/`: `sepsis_fold0_*.csv` and `sepsis_loco_*.csv` (method `tools/sepsis_states.py`, rules `tools/states.json`).
+- Derived analyses (`within_*`, `cross_*`), the confound audit (`confound_*`) and two verifier passes (`verify_recompute_*`, `verify_confound_*`), all in `out/scratch_vs_lora_20261005/`.
+
+The synthesis ran on CPU only and changed no existing file.
+
+## 1. Question
+
+Given the same training pairs, how well does a model trained from scratch generate rare cell states? "From scratch" means random initialisation with no corpus pretraining. The comparison is the pretrained model fine-tuned with LoRA or with full fine-tuning.
+
+## 2. Data, metric and arms
+
+**Metric.**
+- For each arm and state, ratio = generated state % of parent / observed state % of parent.
+- Observed cells are binomially thinned to the arm's generated parent median library, with one constant per arm and marker state. The two scANVI-label states are not thinned.
+- A ratio of 1.00 means the observed frequency. Fidelity is |log2 ratio|.
+- For a scratch-pretrained pair, D = fidelity(scratch) - fidelity(pretrained). D > 0 means pretrained is closer.
+- Brackets are 95% donor-bootstrap intervals (2,000 resamples, seed 20261005). "BC" is the bias-corrected sensitivity interval.
+- 15 states: 13 rare states, one positive control (ISG-high monocyte) and one negative control (platelet signal in monocytes).
+- All 15 are informative in both settings; the smallest observed-thinned count is 117 within the cohort and 629 across cohorts.
+- No arm generated zero cells of any state, so there are no complete misses.
+
+**Within the cohort.** Fold 0, with 17 held-out donors: 9 adult and 3 paediatric sepsis, 2 adult and 1 paediatric healthy control, and 1 adult and 1 paediatric ICU control. Each arm generated 1,500 cells per donor from real bulk, K = 1,000.
+
+| Arm | Initialisation | Pairs | Training time |
+|---|---|---|---|
+| zs__raw | pretrained, zero-shot | 0 | none |
+| lora_n5__raw | pretrained + LoRA | 5 | 3.2 min |
+| scratch_n5__raw | scratch | 5 | 61 min |
+| real_lora__raw | pretrained + LoRA | 58 | 5.4 min |
+| real_full__raw | pretrained + full FT | 58 | 4.7 min |
+| scratch_full__raw | scratch | 58 | 79 min (100k-step cap) |
+| lora_recipe__raw | pretrained + LoRA + recipes | 58 | 4.5 min |
+| scratch_recipe__raw | scratch + recipes | 58 | 86 min |
+
+Scratch needs 15-19x the training time of the matched pretrained arm.
+
+**Across cohorts.** All 82 usable sepsis donors were held out: 46 adult and 15 paediatric sepsis, 7 adult and 4 paediatric healthy controls, and 6 adult and 4 paediatric ICU controls. Models were trained on ImMicro, Imperial, Giroux and Liu pairs only. Each arm generated 1,500 cells per donor.
+
+| Arm | Initialisation / method | Training time |
+|---|---|---|
+| zs__raw | pretrained, zero-shot | none |
+| xcalL_K1000 | pretrained zero-shot + cross-cohort calibration | none |
+| locoL_lora16sat_s26750_K1000 | pretrained + LoRA r16, step-saturated | 20 min |
+| locoL_lora64sat_s59000_K1000 | pretrained + LoRA r64 | 44 min |
+| locoL_fullsat_s108500_K1000 (and s20000) | pretrained + full FT, saturated at 108.5k steps | 93 min |
+| locoL_scratchsat_s162000_K1000 (milestones s50000/s100000/s150000) | scratch; s162000 is the selected final state (step-152,000 weights, chosen by inner-validation loss) | 127 min |
+| real_lora__raw, lora_n5__raw | own-cohort references: pretrained LoRA on the sepsis cohort's own pairs, cross-fitted 5-fold (~58 or 5 pairs per fold) | fold 0: 5.4 / 3.2 min |
+
+Scratch needs 1.4-6.4x the training time of the pretrained cross-cohort arms.
+
+**Benchmark context.** These numbers come from the paired cross-cohort benchmark and were not re-derived here.
+- The cross-cohort scratch model has precision 1.00 but coverage 0.03-0.06. LoRA r16 has coverage 0.34 and own-cohort LoRA 0.83.
+- Within the cohort, scratch matched pretrained on coverage and precision. With 5 pairs, however, it lost donor-specific composition: across-donor composition r was 0.17 vs 0.70.
+
+**Verdict rule** (both result tables).
+- **pretrained** or **scratch**: that initialisation is closer in most matched comparisons. At least one difference is significant (percentile or BC interval), none is significant the other way by the percentile interval, and the ordering survives the depth and composition checks.
+- **both-fail**: scratch and every pretrained fine-tuned comparator are at 0.30x or less, so any ordering compares two failures.
+- **tie**: everything else.
+
+The matched comparisons are:
+- Within the cohort: 5 pairs, 58 pairs vs LoRA, 58 pairs vs full FT, and 58 pairs with recipes.
+- Across cohorts: final scratch vs LoRA r16, LoRA r64, full FT 108.5k and xcal.
+
+## 3. Within-cohort result (same pairs, fold 0)
+
+Values are generated/observed ratios; 1.00 is the observed frequency.
+
+| State | zero-shot | LoRA-5 | scratch-5 | LoRA-58 | full FT-58 | scratch-58 | LoRA+recipe-58 | scratch+recipe-58 | Verdict |
+|---|---|---|---|---|---|---|---|---|---|
+| C1Q-high monocyte (reference) | 0.02 | 1.07 | 1.01 | 1.05 | 0.92 | 0.47 | 0.63 | 0.53 | pretrained |
+| MS1 monocyte (IL1R2+) | 0.08 | 0.15 | 0.10 | 0.21 | 0.36 | 0.40 | 0.30 | 0.46 | scratch |
+| Cycling monocyte (Ye 2026) | 0.09 | 0.11 | 0.06 | 0.09 | 0.11 | 0.09 | 0.10 | 0.09 | both-fail |
+| Cycling T/NK | 2.52 | 1.15 | 0.51 | 0.93 | 1.34 | 0.42 | 0.99 | 0.48 | pretrained |
+| Cycling activated T (CD38/HLA-DR) | 0.65 | 0.44 | 0.40 | 0.41 | 0.58 | 0.31 | 0.46 | 0.28 | tie |
+| Proliferating T (label) | 3.88 | 2.49 | 1.37 | 1.49 | 1.85 | 1.13 | 1.76 | 0.98 | scratch |
+| ISG-high T | 2.71 | 0.46 | 0.15 | 0.37 | 0.45 | 0.17 | 0.53 | 0.18 | pretrained |
+| Cytotoxic CD4 T | 0.47 | 1.79 | 1.52 | 0.90 | 1.32 | 0.80 | 1.23 | 1.19 | tie |
+| Plasmablast (label) | 1.51 | 1.87 | 1.12 | 3.47 | 3.28 | 1.24 | 2.30 | 1.18 | scratch |
+| Plasmablast (markers) | 0.78 | 0.50 | 0.35 | 1.04 | 1.15 | 0.64 | 0.71 | 0.73 | tie |
+| DN2 / ABC B | 1.27 | 1.11 | 1.63 | 0.35 | 0.37 | 0.63 | 0.28 | 0.46 | scratch |
+| HSPC | 0.88 | 0.96 | 0.48 | 0.57 | 0.78 | 1.34 | 0.71 | 0.96 | tie |
+| Granulocyte-like (LDG) | 0.14 | 0.05 | 0.13 | 0.04 | 0.04 | 0.04 | 0.03 | 0.06 | both-fail |
+| (control) ISG-high monocyte, positive | 2.63 | 0.32 | 0.32 | 0.44 | 0.58 | 0.31 | 0.60 | 0.26 | tie |
+| (control) Platelet signal in monocytes, negative | 0.36 | 0.84 | 0.60 | 0.55 | 0.70 | 0.39 | 0.54 | 0.45 | pretrained |
+| Mean abs log2, 13 rare states | 1.76 | 1.30 | 1.52 | 1.39 | 1.27 | 1.41 | 1.39 | 1.31 | |
+
+### Per-state notes
+
+**C1Q-high (pretrained, weak).**
+- At 5 pairs both reach the observed level (1.01x vs 1.07x). Scratch-5, however, is donor-collapsed and over-generates CD16 monocytes; composition-standardised, the ratios are 0.61x vs 0.85x.
+- At 58 pairs scratch falls to 0.47x, against LoRA 1.05x (D +1.01 [-0.26, 1.34]) and full FT 0.92x (D +0.97 [-0.10, 1.25]).
+- These differences are significant only by BC and after composition standardisation (+0.82 [0.24, 1.20]; +0.83 [0.35, 1.33]). They are robust to the depth scheme.
+- Fold 0 is this LoRA's best C1Q fold: 1.04x vs 0.27-0.76x in folds 1-4, and 0.50x pooled. A fold-0 LoRA trained on 40 pairs gives only 0.44x.
+
+**MS1 (scratch, weak).**
+- Both are near 0.1x at 5 pairs.
+- At 58 pairs scratch is 0.40x vs LoRA 0.21x (D -0.97 [-1.44, 0.04]); this is BC-significant in every depth scheme.
+- Scratch ties full FT (0.36x). With recipes it is 0.46x vs 0.30x, n.s.
+- Fold 0 is this LoRA's worst MS1 fold: 0.72-1.21x elsewhere, 0.71x pooled.
+
+**Cycling monocyte (both-fail).** 0.06-0.11x in every trained arm. No difference is significant.
+
+**Cycling T/NK (pretrained).**
+- Pretrained 0.93-1.34x vs scratch 0.42-0.51x. D > 0 in 4 of 4 comparisons.
+- Significant vs 58-pair LoRA (+1.15 [0.20, 1.78]), and still significant after T-subtype composition standardisation (+1.00 [0.15, 1.59]).
+- Fold 0 has the lowest observed level of the five folds. The same LoRA gives 0.34-0.66x in the other folds.
+
+**Cycling activated T (tie).**
+- Every arm under-generates (0.28-0.58x).
+- Under the screen's scoring, pretrained is closer and significantly so vs full FT (+0.87 [0.24, 1.40]).
+- However, 82-94% of the pretrained arms' cycling-rule T cells sit inside their over-generated proliferating-T label (observed 72-74%).
+- After composition standardisation, scratch is significantly closer at 5 pairs (-0.59 [-1.03, -0.17]) and the full-FT advantage is gone (+0.36 [-0.14, 0.75]).
+
+**Proliferating T, label (scratch, specific to fold 0).**
+- The pretrained models over-generate this label: 1.49-2.49x vs 0.98-1.37x for scratch. Scratch is significantly closer in 3 of 4 comparisons.
+- This is not better biology. Fewer scratch label cells pass the cycling-marker rule (16-23% vs 23-28%; observed 36-37%). That gap is significant only at 58 pairs.
+- In folds 1-4 the same 58-pair LoRA under-generates the label (0.49-1.09x; 0.78x pooled).
+
+**ISG-high T (pretrained).**
+- Pretrained 0.37-0.53x vs scratch 0.15-0.18x. D > 0 in 4 of 4; significant with recipes (+1.52 [0.08, 2.49]).
+- The ordering holds in every depth scheme and after composition standardisation.
+- Both under-generate. Scratch also under-generates the ISG positive control and loses its disease directions, so this is a general interferon gap in scratch.
+
+**Cytotoxic CD4 T (tie).**
+- 0.80-1.79x in every trained arm, with no fidelity difference.
+- Scratch-5 loses donor tracking (Spearman -0.40 vs 0.74) and reverses two contrasts.
+
+**Plasmablast, label (scratch, specific to fold 0).**
+- Pretrained over-generates (1.87-3.47x vs 1.12-1.24x) because it makes too many plasma cells: 1.25-1.89% of cells vs 0.64% observed, against 0.61-0.67% for scratch.
+- The difference is significant only vs full FT (-1.40 [-2.02, -0.13]), and that significance is lost under donor-mean weighting.
+- Pooled over all 82 donors, the 58-pair LoRA gives 0.76x.
+
+**Plasmablast, markers (tie).**
+- Under the screen's scoring, pretrained is closer at 58 pairs (1.04-1.15x vs 0.64x).
+- Once depth is matched, pretrained overshoots (1.49-1.90x) and scratch becomes closer.
+- At 5 pairs pretrained is robustly closer.
+
+**DN2 / ABC B (scratch at 58 pairs).**
+- At 58 pairs scratch is 0.46-0.63x vs pretrained 0.28-0.37x. This is robust in every depth scheme and significant with recipes (-0.70 [-1.21, -0.08]).
+- At 5 pairs LoRA is closer (BC only), and scratch-5 strongly reverses the adult contrast.
+
+**HSPC (tie).** Each arm generated only 13-35 cells, so the intervals are wide.
+
+**LDG (both-fail).** 0.03-0.13x in every trained arm. The two significant "scratch closer" calls compare two failures.
+
+**Controls.**
+- ISG-high monocyte: no significant difference.
+- Platelet signal: pretrained is closer in 4 of 4, significantly vs full FT (+0.83 [0.45, 1.19]). This is a technical signal, so being closer is not a biological win.
+
+### Aggregate
+
+The pre-specified 13-state mean shows no significant difference:
+- 5 pairs: D +0.22 [-0.08, 0.41] (BC [0.01, 0.46]).
+- 58 pairs: +0.02 [-0.35, 0.25] vs LoRA and +0.14 [-0.14, 0.41] vs full FT.
+- Recipes: -0.09 [-0.38, 0.18].
+- State-level tests: 3 of 52 favour pretrained, 7 favour scratch and 42 are n.s.; with BC intervals it is 9 vs 11.
+
+This is weaker than "equally faithful", for four reasons:
+- **Non-significance, not equivalence.** The intervals span about +-0.3 log2.
+- **The mean hides a T/NK advantage for pretrained.** Over the five T/NK states, pretrained is significantly closer at 58 pairs:
+  - vs LoRA +0.48 [0.07, 0.88], significant in all 7 scoring variants;
+  - vs full FT +0.45 [0.01, 1.02];
+  - with recipes +0.48 [0.01, 0.91].
+  Over the 11 marker states, full FT is significantly closer (+0.35 [0.02, 0.62]) in 4 of 7 variants.
+- **The 5-pair BC result is a depth/composition artifact.** At common depth D is +0.10 [-0.23, 0.35]; after composition standardisation it is -0.00.
+- **Fold 0 is LoRA-5's best fold.** Its mean abs log2 is 1.26-1.30 here vs 1.53-2.94 in folds 1-4.
+
+Zero-shot (1.76) is the least faithful arm. The two 58-pair scratch arms beat it (D -0.35 [-0.66, -0.13] and -0.46 [-0.73, -0.24]), and the gain comes mostly from C1Q-high. Scratch-5 does not beat it significantly (D -0.24 [-0.52, 0.01]).
+
+**Donor tracking.** Mean Spearman, pretrained vs scratch:
+
+| Comparison | Pretrained | Scratch | Difference |
+|---|---|---|---|
+| 5 pairs | 0.30 | 0.19 | n.s. |
+| LoRA-58 | 0.29 | 0.27 | n.s. |
+| Full FT-58 | 0.43 | 0.27 | +0.15, BC [0.003, 0.33] |
+| Recipes | 0.40 | 0.36 | n.s. |
+
+No difference is detectable, both track poorly, and there is a small tilt towards pretrained. A perfect generator would reach 0.78-0.85.
+
+## 4. Cross-cohort result (no sepsis pairs in training)
+
+Values are generated/observed ratios. The column "scratch 162k at r16 depth" scores the same scratch cells against observed cells thinned for LoRA r16.
+
+| State | own LoRA-58 | own LoRA-5 | zero-shot | xcal | LoRA r16 | LoRA r64 | full FT 108.5k | scratch 162k | scratch 162k at r16 depth | scratch 50k | Verdict |
+|---|---|---|---|---|---|---|---|---|---|---|---|
+| C1Q-high monocyte (reference) | 0.50 | 0.53 | 0.03 | 0.07 | 0.03 | 0.06 | 0.09 | 0.07 | 0.08 | 0.09 | both-fail |
+| MS1 monocyte (IL1R2+) | 0.71 | 0.64 | 0.08 | 0.72 | 0.12 | 0.11 | 0.27 | 0.10 | 0.16 | 0.16 | both-fail |
+| Cycling monocyte (Ye 2026) | 0.23 | 0.30 | 0.10 | 0.25 | 0.06 | 0.07 | 0.11 | 0.02 | 0.04 | 0.06 | both-fail |
+| Cycling T/NK | 0.54 | 0.29 | 0.83 | 0.46 | 0.15 | 0.17 | 0.36 | 0.05 | 0.06 | 0.06 | pretrained |
+| Cycling activated T (CD38/HLA-DR) | 0.30 | 0.16 | 0.36 | 0.25 | 0.10 | 0.14 | 0.23 | 0.05 | 0.07 | 0.08 | both-fail |
+| Proliferating T (label) | 0.78 | 0.62 | 1.96 | 0.64 | 0.79 | 1.04 | 1.13 | 0.53 | 0.53 | 0.44 | pretrained |
+| ISG-high T | 0.57 | 0.39 | 1.74 | 1.59 | 1.66 | 1.93 | 2.05 | 0.60 | 1.00 | 0.70 | tie |
+| Cytotoxic CD4 T | 1.35 | 1.06 | 0.43 | 2.02 | 1.41 | 1.70 | 1.61 | 0.94 | 1.07 | 0.86 | scratch |
+| Plasmablast (label) | 0.76 | 0.41 | 0.39 | 0.78 | 1.38 | 1.01 | 2.21 | 0.87 | 0.87 | 0.49 | scratch |
+| Plasmablast (markers) | 0.33 | 0.19 | 0.28 | 0.37 | 0.78 | 0.69 | 1.47 | 0.34 | 0.53 | 0.26 | pretrained |
+| DN2 / ABC B | 0.31 | 0.42 | 1.49 | 0.30 | 0.17 | 0.18 | 0.28 | 0.16 | 0.24 | 0.16 | both-fail |
+| HSPC | 0.40 | 0.71 | 0.55 | 1.68 | 1.62 | 2.96 | 2.97 | 1.31 | 1.47 | 0.96 | scratch |
+| Granulocyte-like (LDG) | 0.07 | 0.04 | 0.24 | 0.06 | 0.01 | 0.02 | 0.02 | 0.01 | 0.02 | 0.04 | both-fail |
+| (control) ISG-high monocyte, positive | 0.61 | 0.55 | 1.80 | 1.42 | 1.31 | 1.45 | 1.77 | 0.47 | 0.95 | 0.53 | tie |
+| (control) Platelet signal in monocytes, negative | 0.56 | 0.57 | 0.28 | 0.19 | 0.18 | 0.18 | 0.36 | 0.08 | 0.17 | 0.13 | pretrained |
+| Mean abs log2, 13 rare states | 1.29 | 1.54 | 1.82 | 1.53 | 2.32 | 2.21 | 1.92 | 2.62 | 2.21 | 2.33 | |
+
+### Per-state notes
+
+**C1Q-high (both-fail).** Every cross-cohort arm is at 0.03-0.09x, while own-cohort LoRA reaches 0.50-0.53x. Scratch (0.07x) is significantly closer than r16 (0.03x), but both fail.
+
+**MS1 (both-fail).**
+- Scratch (0.10x), r16 (0.12x) and r64 (0.11x) do not differ; after marker-depth matching it is 0.14x vs 0.13x.
+- Full FT is modestly but robustly higher (0.27x).
+- xcal reaches 0.72x [0.55, 1.05] with no sepsis pairs and no training. That is as high as own-cohort LoRA.
+
+**Cycling monocyte (both-fail).**
+- 0.02-0.30x in every arm, own-cohort LoRA included.
+- Scratch is lowest: 21 of 20,808 monocytes.
+- At matched depth, pretrained stays closer only vs full FT and xcal.
+
+**Cycling T/NK (pretrained; confirmed).**
+- Scratch makes 35 cycling cells among 72,018 T/NK cells (0.05x), vs 0.15-0.46x in every pretrained arm.
+- Pretrained is significantly closer in all 4 comparisons. This survives matched depth (vs r16, D +1.23 [0.61, 1.91]), detection-matched depth and every leave-one-donor-out run.
+
+**Cycling activated T (both-fail; pretrained less bad).** Scratch is at 0.05x vs 0.10-0.25x for the pretrained arms. At matched depth, the pretrained arms are significantly closer vs r64, full FT and xcal.
+
+**Proliferating T, label (pretrained, weak).**
+- Scratch 0.53x vs r16 0.79x (D +0.57 [0.09, 0.80]). Differences vs r64, full FT and xcal are n.s.
+- About 5% of label cells carry 2+ cycling genes in scratch or r16, vs 37% of observed cells. This is a label-frequency result, not cycling evidence.
+
+**ISG-high T (tie).**
+- Pretrained over-generates (1.35-2.05x) and scratch is at 0.60x.
+- At r16 depth scratch is 1.00x; detection-matched it is 0.67x.
+- All differences are n.s.
+
+**Cytotoxic CD4 T (scratch).**
+- Scratch (0.94x) is significantly closer than xcal (2.02x), and at matched depth closer than r64 (1.70x) and full FT (1.61x).
+- But scratch, r16, r64 and full FT all reverse the adult and all-vs-all contrasts.
+- Scratch also tracks donors worse than full FT (Spearman 0.40 vs 0.62).
+
+**Plasmablast, label (scratch, weak).**
+- Scratch (0.87x) is as close as r64 (1.01x).
+- It is significantly closer only than the over-generating full FT (2.21x), and that call holds in only 61 of 82 leave-one-donor-out runs.
+
+**Plasmablast, markers (pretrained).**
+- Scratch 0.34x vs r16 0.78x, r64 0.69x and full FT 1.47x. This survives marker-depth matching (0.47x vs 0.58-1.21x).
+- xcal (0.37x) is no better than scratch, and own-cohort LoRA-58 is also low (0.33x).
+
+**DN2 / ABC B (both-fail).** All cross-cohort arms are at 0.16-0.30x (own-cohort 0.31-0.42x). The ordering flips with the depth reference.
+
+**HSPC (scratch).**
+- Scratch (1.31x) is significantly closer than r64 and full FT (both about 2.96x) at both depth references.
+- The difference is n.s. vs r16 (1.62x) and xcal (1.68x).
+- It depends on the checkpoint: s100000 was at 2.58x.
+
+**LDG (both-fail).** Fine-tuned cross-cohort arms are at 0.01-0.02x; scratch generated 35 of 123,000 cells.
+
+**Controls.**
+- ISG-high monocyte: n.s. Scratch is 0.95x at r16 depth.
+- Platelet signal: pretrained is closer, robust to depth matching. This is a technical signal.
+
+### Aggregate (confirmed)
+
+Under the screen's scoring, scratch is further from observed than every pretrained cross-cohort arm:
+
+| Comparison | D, screen scoring | D, matched depth |
+|---|---|---|
+| vs LoRA r16 | +0.30 [0.09, 0.52] | -0.11 [-0.24, 0.11] |
+| vs LoRA r64 | +0.41 [0.17, 0.62] | -0.04 [-0.21, 0.13] |
+| vs full FT | +0.69 [0.42, 1.00] | +0.25 [0.03, 0.54] |
+| vs xcal | +1.08 [0.92, 1.29] | +0.68 [0.55, 0.89] |
+
+- At matched depth the gap to LoRA vanishes.
+- The deficit vs full FT is marginal and carried by the cycling and MS1 states.
+- xcal has the lowest mean error of any cross-cohort arm (1.53, vs 1.75-2.62 for the others).
+
+### Scratch training length
+
+The final checkpoint is further from observed than the 50k checkpoint: D +0.29 [0.15, 0.43].
+- Longer training moved C1Q-high, MS1, cycling monocytes, cycling activated T and LDG further away.
+- It improved only the two plasmablast states and proliferating T (label).
+- At matched depth, s50000 beats r16 (D -0.39 [-0.54, -0.17]) while s162000 only ties it.
+- Over the same steps the library inflated from 13,891 to 20,606 UMIs and coverage fell from 0.063 to 0.032. This is a co-trend across four checkpoints, not a tested cause.
+
+### Donor tracking
+
+- No state differs between scratch and r16 or r64 (0 of 13 each), but this is a low-power null.
+- Full FT and xcal each track 2 states better than scratch.
+- Own-cohort LoRA-58 tracks best on average (mean rho 0.36), although xcal tracks MS1 better.
+
+## 5. Contrasts
+
+### Within the cohort (anecdotal: 2 adult and 1 paediatric healthy control)
+
+Directions kept, out of 25 eligible state-contrast pairs:
+
+| Setting | Pretrained | Scratch |
+|---|---|---|
+| 5 pairs | LoRA-5: 23 | 17 |
+| 58 pairs | LoRA-58: 20; full FT-58: 20 | 19 |
+| 58 pairs + recipes | 23 | 16 |
+
+Zero-shot keeps 21.
+
+- Only the recipe gap is significant: 7 vs 0 discordant pairs, p = 0.016, uncorrected and not independent.
+- At 58 pairs the gap is 2 vs 1 and 3 vs 2 discordant pairs (p = 1).
+- At 5 pairs, scratch has 3 strong reversals: DN2 adult (+4.74 vs observed -3.16 pp) and cytotoxic CD4 T adult and all-sepsis. LoRA-5 has 2.
+- Eight of the 25 pairs rest on a single paediatric control.
+
+### Across cohorts (control groups of 7, 4 and 21)
+
+Directions kept, out of 21 common pairs:
+
+| Arm | Kept |
+|---|---|
+| Scratch | 14 |
+| LoRA r16 | 11 |
+| LoRA r64 | 10 |
+| Full FT | 16 |
+| xcal | 19 |
+| Own-cohort LoRA-58 | 19 |
+
+Scratch, r16 and r64 are not distinguishable from coin flipping.
+
+Kept effects are compressed. The share of the observed difference that the median kept pair reproduces is:
+
+| Arm | Share reproduced |
+|---|---|
+| Scratch | 4% (7% at r16 depth) |
+| LoRA and full FT | 13-15% |
+| xcal | 23% |
+| Own-cohort LoRA-58 | 34% |
+
+- On pairs kept by both arms, at matched depth, scratch is only about 1.3x more compressed than LoRA (n.s.).
+- Compression relative to full FT, xcal and own-cohort LoRA is robust.
+- MS1, adult sepsis vs healthy control: scratch reproduces 6%, r16 10%, full FT 25%, xcal 50% and own-cohort LoRA-58 62%.
+- C1Q-high: scratch reproduces 1-4%, r16 reverses all three contrasts, and own-cohort LoRA-58 reproduces 30-34%.
+
+## 6. Confounds and caveats
+
+1. **Cross-cohort depth confound (high).**
+   - Scratch cells carry 20,606-UMI libraries but only 236 detected genes, and about 0.1% of their counts fall on marker genes. Their marker detection matches the ~1,000-UMI pretrained arms.
+   - The screen compared scratch with unthinned observed cells, so its marker-state ratios are biased down by 1.1-2.0x (up to 3.2x with detection matching).
+   - Library thinning is also imperfect for the pretrained arms. Every absolute cross-cohort ratio therefore carries a shared bias.
+2. **V1 thinning is invalid for cross-cohort scratch.** It removes ~95% of marker counts. Its "complete misses" are artifacts and should not be quoted.
+3. **Cross-cohort scratch count profiles are unrealistic.** Libraries inflate, cell-to-cell spread is compressed (q90/q10 2.0 vs 7.3 observed) and coverage is 0.03. The full generated matrices were not examined.
+4. **Scratch-5 is donor-collapsed** (composition r 0.17 vs 0.70). Its pooled ratios overstate donor-level fidelity.
+5. **One fold, one model and one seed within the cohort.**
+   - Fold 0 favours LoRA-5, and LoRA-58 on C1Q-high, but disfavours LoRA-58 on MS1.
+   - It also has the lowest observed cycling-T and plasmablast levels.
+   - Scratch exists only for fold 0.
+6. **Label states depend on scANVI calls.** The pretrained arms over-generate plasma cells and proliferating-labelled T cells.
+7. **Controls.**
+   - Scratch's ISG deficit reflects a general interferon gap.
+   - "Closer to observed" also rewards technical signal, as the platelet control shows.
+8. **Zero-shot ISG overshoot is mostly depth.** ISG-high T goes from 2.71x to 0.76x after matching.
+9. **Statistics.**
+   - 17 donors within the cohort.
+   - 104 state-level tests with no multiplicity correction.
+   - Within-cohort contrasts rest on 1-2 controls.
+10. **Checkpoints.** The scratch checkpoint was chosen by validation loss, and the 50k checkpoint is closer. The within-cohort 58-pair scratch run hit its 100k-step cap.
+
+### Statements dropped after verification
+
+- **"Pairs on the pretrained model, not pretraining alone, recover the monocyte states."** False for MS1, since xcal reaches 0.72x. The pretrained model is also not shown to be necessary: within the cohort, scratch with 58 pairs reaches C1Q 0.47x and MS1 0.40-0.46x.
+- **"Scratch compression is 3-4x worse than LoRA."** It is about 1.3x on shared pairs at matched depth, and n.s.
+- **"Only plasmablasts improve with longer scratch training."** Proliferating T (label) improves too.
+- **"The only significant scratch wins are HSPC, cytotoxic CD4 T and plasmablast label."** Plasmablast markers vs xcal at matched depth is another.
+- **"Zero-shot is closer for cycling activated T."** This is a depth/composition artifact.
+- **"Pretrained is better on T cycling and interferon states in every comparison."** This fails for cycling activated T after composition standardisation.
+- **"Donor tracking is similar."** The data support only "no detectable difference".
+- **"Corrected ISG states are about 1.0x."** True only at the r16 reference.
+
+## 7. Verdict
+
+**With the same target-cohort pairs, scratch matches pretrained on the pre-specified 13-state average, but not everywhere.**
+- The difference on the average is not significant at 5 or 58 pairs, and equivalence is not shown.
+- Pretrained is better on proliferating and interferon T/NK states: cycling T/NK and ISG-high T, with the T/NK block significantly closer at 58 pairs.
+- Scratch collapses donors with 5 pairs and needs 15-19x the training time.
+- Scratch is closer mainly where the pretrained models over-call fold-0 scANVI labels, and on DN2 and the fold-specific MS1 result at 58 pairs.
+- Both fail LDG and cycling monocytes.
+
+**Without target-cohort pairs, initialisation is not the main factor.**
+- Every cross-cohort fine-tuned model, scratch or pretrained, fails the sepsis monocyte states: C1Q-high at 0.09x or less and MS1 at 0.27x or less. Own-cohort LoRA reaches about 0.5x and 0.7x.
+- After the depth correction, scratch ties LoRA on the aggregate.
+- Scratch is clearly worse on cycling T/NK (0.05x vs 0.15-0.46x; confirmed). It also trails full FT and xcal; xcal has the lowest error of any cross-cohort arm and brings MS1 to 0.72x.
+- Scratch is closer only where the pretrained arms over-generate (HSPC, cytotoxic CD4 T).
+- The scratch generator drifts with longer training: libraries inflate, coverage reaches 0.03 and the monocyte states get worse.
+
+**Bottom line.**
+- Pretraining alone does not produce the sepsis monocyte states: zero-shot gives C1Q-high 0.03x and MS1 0.08x.
+- Within the cohort, its own pairs recover C1Q-high under either initialisation.
+- Across cohorts, only own-cohort pairs recover C1Q-high. MS1 is also recovered by calibrating the pretrained model.
+- For rare states, pretraining adds four things:
+  - better proliferating T/NK states;
+  - donor-specific composition with few pairs;
+  - 15-19x shorter training;
+  - when there are no target pairs, a training-free calibrated model with the lowest average error of any cross-cohort arm.
+
+**Open points.** No new runs were made. These remain untested:
+- scratch on folds 1-4 and with a second seed;
+- scratch checkpoint selection by rare-state fidelity;
+- scratch combined with the xcal calibration.
+
+**Archive.** No pbisc-diffusion delta was published by this subagent.

paper-releases/pbisc-diffusion/updates/20261005T142225Z-rare-states-scratch-vs-pretrained/diffusion/dataset/server_1/model8/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/confound_collapse_composition.csv

@@ -0,0 +1,121 @@
+setting,arm,arm_class,cell_type,sd_gen_pct,sd_obs_pct,sd_ratio_gen_over_obs,pearson_donors,spearman_donors,mean_gen_pct,mean_obs_pct
+fold0,zs__raw,pretrained zero-shot,Monocyte,22.1062,18.704,1.18189,0.508468,0.482843,37.2784,24.6001
+fold0,zs__raw,pretrained zero-shot,T cell,14.6595,18.3795,0.7976,0.560715,0.566176,33.8471,44.4687
+fold0,zs__raw,pretrained zero-shot,B cell,13.4188,10.237,1.31082,0.638376,0.517157,17.3529,16.693
+fold0,zs__raw,pretrained zero-shot,NK cell,3.76526,9.24776,0.407154,0.732567,0.470877,7.77647,8.56252
+fold0,zs__raw,pretrained zero-shot,Plasma cell (L2),1.53872,0.612497,2.5122,0.550451,0.161506,0.819608,0.639072
+fold0,zs__raw,pretrained zero-shot,MEAN over 4 L1 types,null,null,null,0.610031,null,null,null
+fold0,lora_n5__raw,pretrained LoRA (5 pairs),Monocyte,14.7342,18.704,0.787757,0.594242,0.588235,28.1686,24.6001
+fold0,lora_n5__raw,pretrained LoRA (5 pairs),T cell,14.1944,18.3795,0.772296,0.656026,0.661765,48.6118,44.4687
+fold0,lora_n5__raw,pretrained LoRA (5 pairs),B cell,8.65585,10.237,0.845549,0.765117,0.683824,15.6745,16.693
+fold0,lora_n5__raw,pretrained LoRA (5 pairs),NK cell,2.02399,9.24776,0.218863,0.788151,0.513199,3.78039,8.56252
+fold0,lora_n5__raw,pretrained LoRA (5 pairs),Plasma cell (L2),0.924892,0.612497,1.51004,0.570121,0.275865,1.01961,0.639072
+fold0,lora_n5__raw,pretrained LoRA (5 pairs),MEAN over 4 L1 types,null,null,null,0.700884,null,null,null
+fold0,scratch_n5__raw,scratch (5 pairs),Monocyte,8.29876,18.704,0.443689,0.359853,0.441176,24.4745,24.6001
+fold0,scratch_n5__raw,scratch (5 pairs),T cell,8.04542,18.3795,0.43774,0.577427,0.521153,53.3765,44.4687
+fold0,scratch_n5__raw,scratch (5 pairs),B cell,1.13682,10.237,0.11105,0.191178,-0.0838487,11.4118,16.693
+fold0,scratch_n5__raw,scratch (5 pairs),NK cell,1.23512,9.24776,0.133559,-0.444709,-0.477248,8.02353,8.56252
+fold0,scratch_n5__raw,scratch (5 pairs),Plasma cell (L2),0.240642,0.612497,0.392887,0.653691,0.38619,0.611765,0.639072
+fold0,scratch_n5__raw,scratch (5 pairs),MEAN over 4 L1 types,null,null,null,0.170937,null,null,null
+fold0,real_lora__raw,pretrained LoRA (58 pairs),Monocyte,10.6957,18.704,0.571841,0.668748,0.687922,15.2588,24.6001
+fold0,real_lora__raw,pretrained LoRA (58 pairs),T cell,12.3678,18.3795,0.672916,0.840812,0.82598,55.4784,44.4687
+fold0,real_lora__raw,pretrained LoRA (58 pairs),B cell,12.6267,10.237,1.23345,0.771967,0.686275,19.2275,16.693
+fold0,real_lora__raw,pretrained LoRA (58 pairs),NK cell,4.33001,9.24776,0.468222,0.80124,0.767627,5.89804,8.56252
+fold0,real_lora__raw,pretrained LoRA (58 pairs),Plasma cell (L2),4.19079,0.612497,6.84215,0.490924,0.0135386,1.8902,0.639072
+fold0,real_lora__raw,pretrained LoRA (58 pairs),MEAN over 4 L1 types,null,null,null,0.770692,null,null,null
+fold0,real_full__raw,pretrained full FT (58 pairs),Monocyte,8.6191,18.704,0.460816,0.475826,0.416667,16.1569,24.6001
+fold0,real_full__raw,pretrained full FT (58 pairs),T cell,9.68523,18.3795,0.526959,0.750745,0.740196,52.6863,44.4687
+fold0,real_full__raw,pretrained full FT (58 pairs),B cell,13.1472,10.237,1.28429,0.816516,0.698529,20.5216,16.693
+fold0,real_full__raw,pretrained full FT (58 pairs),NK cell,3.47573,9.24776,0.375845,0.967568,0.843137,5.41176,8.56252
+fold0,real_full__raw,pretrained full FT (58 pairs),Plasma cell (L2),2.85682,0.612497,4.66422,0.541464,0.373006,1.78431,0.639072
+fold0,real_full__raw,pretrained full FT (58 pairs),MEAN over 4 L1 types,null,null,null,0.752664,null,null,null
+fold0,scratch_full__raw,"scratch (58 pairs, 100k cap)",Monocyte,8.65503,18.704,0.462736,0.522557,0.490196,15.8275,24.6001
+fold0,scratch_full__raw,"scratch (58 pairs, 100k cap)",T cell,10.5052,18.3795,0.571572,0.653635,0.591049,52.7176,44.4687
+fold0,scratch_full__raw,"scratch (58 pairs, 100k cap)",B cell,11.0098,10.237,1.07549,0.914161,0.762255,19.8157,16.693
+fold0,scratch_full__raw,"scratch (58 pairs, 100k cap)",NK cell,3.63582,9.24776,0.393157,0.627379,0.685469,8.45098,8.56252
+fold0,scratch_full__raw,"scratch (58 pairs, 100k cap)",Plasma cell (L2),0.778489,0.612497,1.27101,0.630419,0.432004,0.67451,0.639072
+fold0,scratch_full__raw,"scratch (58 pairs, 100k cap)",MEAN over 4 L1 types,null,null,null,0.679433,null,null,null
+fold0,lora_recipe__raw,pretrained LoRA + recipes (58 pairs),Monocyte,9.32671,18.704,0.498648,0.620439,0.580882,15.549,24.6001
+fold0,lora_recipe__raw,pretrained LoRA + recipes (58 pairs),T cell,10.8014,18.3795,0.587688,0.805984,0.817903,54.3176,44.4687
+fold0,lora_recipe__raw,pretrained LoRA + recipes (58 pairs),B cell,12.8602,10.237,1.25625,0.844425,0.688725,20.2235,16.693
+fold0,lora_recipe__raw,pretrained LoRA + recipes (58 pairs),NK cell,3.1026,9.24776,0.335498,0.948397,0.797057,5.29412,8.56252
+fold0,lora_recipe__raw,pretrained LoRA + recipes (58 pairs),Plasma cell (L2),2.10139,0.612497,3.43086,0.58561,0.131531,1.25098,0.639072
+fold0,lora_recipe__raw,pretrained LoRA + recipes (58 pairs),MEAN over 4 L1 types,null,null,null,0.804811,null,null,null
+fold0,scratch_recipe__raw,scratch + recipes (58 pairs),Monocyte,10.2401,18.704,0.547481,0.727355,0.692827,14.1098,24.6001
+fold0,scratch_recipe__raw,scratch + recipes (58 pairs),T cell,11.9675,18.3795,0.651135,0.74664,0.671569,54.1765,44.4687
+fold0,scratch_recipe__raw,scratch + recipes (58 pairs),B cell,12.1732,10.237,1.18914,0.897592,0.898835,19.8078,16.693
+fold0,scratch_recipe__raw,scratch + recipes (58 pairs),NK cell,4.71564,9.24776,0.509923,0.705362,0.713235,8.56471,8.56252
+fold0,scratch_recipe__raw,scratch + recipes (58 pairs),Plasma cell (L2),0.614211,0.612497,1.0028,0.530506,0.16718,0.639216,0.639072
+fold0,scratch_recipe__raw,scratch + recipes (58 pairs),MEAN over 4 L1 types,null,null,null,0.769237,null,null,null
+loco,zs__raw,pretrained zero-shot,Monocyte,22.9811,18.7801,1.2237,0.627516,0.554764,36.9707,20.8171
+loco,zs__raw,pretrained zero-shot,T cell,16.9925,18.6943,0.908971,0.609423,0.622976,35.2415,48.4547
+loco,zs__raw,pretrained zero-shot,B cell,12.2515,15.4945,0.790703,0.608653,0.58834,14.8911,17.2757
+loco,zs__raw,pretrained zero-shot,NK cell,4.62488,7.26453,0.636639,0.308145,0.361408,9.40325,8.07229
+loco,zs__raw,pretrained zero-shot,Plasma cell (L2),1.14679,3.54471,0.323521,0.680618,0.332079,0.55122,1.47858
+loco,zs__raw,pretrained zero-shot,MEAN over 4 L1 types,null,null,null,0.538434,null,null,null
+loco,xcalL_K1000,pretrained zero-shot + cross-cohort calibration,Monocyte,20.9104,18.7801,1.11343,0.62042,0.525131,26.4195,20.8171
+loco,xcalL_K1000,pretrained zero-shot + cross-cohort calibration,T cell,19.0005,18.6943,1.01638,0.628309,0.663964,46.8398,48.4547
+loco,xcalL_K1000,pretrained zero-shot + cross-cohort calibration,B cell,16.8284,15.4945,1.08609,0.688565,0.743872,14.2049,17.2757
+loco,xcalL_K1000,pretrained zero-shot + cross-cohort calibration,NK cell,6.36688,7.26453,0.876434,0.454724,0.421662,9.58862,8.07229
+loco,xcalL_K1000,pretrained zero-shot + cross-cohort calibration,Plasma cell (L2),1.97712,3.54471,0.557766,0.671116,0.413264,1.11301,1.47858
+loco,xcalL_K1000,pretrained zero-shot + cross-cohort calibration,MEAN over 4 L1 types,null,null,null,0.598004,null,null,null
+loco,locoL_lora16sat_s26750_K1000,pretrained LoRA r16 (cross-cohort),Monocyte,13.9467,18.7801,0.742632,0.452136,0.423409,18.3122,20.8171
+loco,locoL_lora16sat_s26750_K1000,pretrained LoRA r16 (cross-cohort),T cell,17.1838,18.6943,0.919204,0.502697,0.493774,41.9398,48.4547
+loco,locoL_lora16sat_s26750_K1000,pretrained LoRA r16 (cross-cohort),B cell,17.5288,15.4945,1.13129,0.751937,0.804051,23.4634,17.2757
+loco,locoL_lora16sat_s26750_K1000,pretrained LoRA r16 (cross-cohort),NK cell,5.90033,7.26453,0.812211,0.501971,0.423811,12.0171,8.07229
+loco,locoL_lora16sat_s26750_K1000,pretrained LoRA r16 (cross-cohort),Plasma cell (L2),3.0765,3.54471,0.867912,0.690244,0.355501,1.96748,1.47858
+loco,locoL_lora16sat_s26750_K1000,pretrained LoRA r16 (cross-cohort),MEAN over 4 L1 types,null,null,null,0.552185,null,null,null
+loco,locoL_lora64sat_s59000_K1000,pretrained LoRA r64 (cross-cohort),Monocyte,12.9521,18.7801,0.689671,0.44767,0.416522,16.3041,20.8171
+loco,locoL_lora64sat_s59000_K1000,pretrained LoRA r64 (cross-cohort),T cell,17.7823,18.6943,0.951217,0.38958,0.389105,38.3171,48.4547
+loco,locoL_lora64sat_s59000_K1000,pretrained LoRA r64 (cross-cohort),B cell,19.2426,15.4945,1.2419,0.687854,0.755041,27.4455,17.2757
+loco,locoL_lora64sat_s59000_K1000,pretrained LoRA r64 (cross-cohort),NK cell,6.649,7.26453,0.915269,0.572739,0.461687,12.8098,8.07229
+loco,locoL_lora64sat_s59000_K1000,pretrained LoRA r64 (cross-cohort),Plasma cell (L2),2.96989,3.54471,0.837837,0.752975,0.375191,1.44309,1.47858
+loco,locoL_lora64sat_s59000_K1000,pretrained LoRA r64 (cross-cohort),MEAN over 4 L1 types,null,null,null,0.524461,null,null,null
+loco,locoL_fullsat_s20000_K1000,pretrained full FT 20k (cross-cohort),Monocyte,16.1946,18.7801,0.862326,0.440277,0.435584,19.5138,20.8171
+loco,locoL_fullsat_s20000_K1000,pretrained full FT 20k (cross-cohort),T cell,18.2581,18.6943,0.976668,0.525738,0.528288,40.2276,48.4547
+loco,locoL_fullsat_s20000_K1000,pretrained full FT 20k (cross-cohort),B cell,19.6965,15.4945,1.2712,0.72981,0.799778,24.5593,17.2757
+loco,locoL_fullsat_s20000_K1000,pretrained full FT 20k (cross-cohort),NK cell,5.92675,7.26453,0.815848,0.520724,0.498307,11.8715,8.07229
+loco,locoL_fullsat_s20000_K1000,pretrained full FT 20k (cross-cohort),Plasma cell (L2),5.02026,3.54471,1.41627,0.730937,0.329924,2.53415,1.47858
+loco,locoL_fullsat_s20000_K1000,pretrained full FT 20k (cross-cohort),MEAN over 4 L1 types,null,null,null,0.554137,null,null,null
+loco,locoL_fullsat_s108500_K1000,pretrained full FT 108.5k saturated (cross-cohort),Monocyte,15.1468,18.7801,0.806538,0.413495,0.426794,15.8488,20.8171
+loco,locoL_fullsat_s108500_K1000,pretrained full FT 108.5k saturated (cross-cohort),T cell,19.2485,18.6943,1.02965,0.405599,0.407656,38.6935,48.4547
+loco,locoL_fullsat_s108500_K1000,pretrained full FT 108.5k saturated (cross-cohort),B cell,21.5089,15.4945,1.38816,0.743579,0.795744,27.5447,17.2757
+loco,locoL_fullsat_s108500_K1000,pretrained full FT 108.5k saturated (cross-cohort),NK cell,8.11876,7.26453,1.11759,0.502686,0.450792,13.6081,8.07229
+loco,locoL_fullsat_s108500_K1000,pretrained full FT 108.5k saturated (cross-cohort),Plasma cell (L2),5.96002,3.54471,1.68139,0.657038,0.227854,3.16341,1.47858
+loco,locoL_fullsat_s108500_K1000,pretrained full FT 108.5k saturated (cross-cohort),MEAN over 4 L1 types,null,null,null,0.51634,null,null,null
+loco,locoL_scratchsat_s50000_K1000,scratch 50k (cross-cohort),Monocyte,13.677,18.7801,0.728273,0.439078,0.390048,18.2309,20.8171
+loco,locoL_scratchsat_s50000_K1000,scratch 50k (cross-cohort),T cell,17.5847,18.6943,0.940648,0.532343,0.565264,44.7285,48.4547
+loco,locoL_scratchsat_s50000_K1000,scratch 50k (cross-cohort),B cell,11.1222,15.4945,0.717815,0.676627,0.697799,18.6463,17.2757
+loco,locoL_scratchsat_s50000_K1000,scratch 50k (cross-cohort),NK cell,5.89602,7.26453,0.811618,0.486221,0.383854,15.6358,8.07229
+loco,locoL_scratchsat_s50000_K1000,scratch 50k (cross-cohort),Plasma cell (L2),1.82577,3.54471,0.51507,0.888562,0.360945,0.695122,1.47858
+loco,locoL_scratchsat_s50000_K1000,scratch 50k (cross-cohort),MEAN over 4 L1 types,null,null,null,0.533567,null,null,null
+loco,locoL_scratchsat_s100000_K1000,scratch 100k (cross-cohort),Monocyte,12.8632,18.7801,0.684935,0.444751,0.38049,18.5919,20.8171
+loco,locoL_scratchsat_s100000_K1000,scratch 100k (cross-cohort),T cell,16.6063,18.6943,0.888308,0.508368,0.533604,43.1268,48.4547
+loco,locoL_scratchsat_s100000_K1000,scratch 100k (cross-cohort),B cell,10.5631,15.4945,0.681736,0.657391,0.672294,19.987,17.2757
+loco,locoL_scratchsat_s100000_K1000,scratch 100k (cross-cohort),NK cell,4.99551,7.26453,0.687658,0.522576,0.457876,14.3341,8.07229
+loco,locoL_scratchsat_s100000_K1000,scratch 100k (cross-cohort),Plasma cell (L2),2.23068,3.54471,0.629299,0.872294,0.292091,0.965041,1.47858
+loco,locoL_scratchsat_s100000_K1000,scratch 100k (cross-cohort),MEAN over 4 L1 types,null,null,null,0.533272,null,null,null
+loco,locoL_scratchsat_s150000_K1000,scratch 150k (cross-cohort),Monocyte,13.9967,18.7801,0.745293,0.439341,0.377908,20.0959,20.8171
+loco,locoL_scratchsat_s150000_K1000,scratch 150k (cross-cohort),T cell,18.0262,18.6943,0.964263,0.503292,0.521204,42.0829,48.4547
+loco,locoL_scratchsat_s150000_K1000,scratch 150k (cross-cohort),B cell,12.3963,15.4945,0.800048,0.592228,0.59812,21.4756,17.2757
+loco,locoL_scratchsat_s150000_K1000,scratch 150k (cross-cohort),NK cell,4.46422,7.26453,0.614523,0.477655,0.441803,13.1846,8.07229
+loco,locoL_scratchsat_s150000_K1000,scratch 150k (cross-cohort),Plasma cell (L2),1.96684,3.54471,0.554868,0.87949,0.286109,1.13415,1.47858
+loco,locoL_scratchsat_s150000_K1000,scratch 150k (cross-cohort),MEAN over 4 L1 types,null,null,null,0.503129,null,null,null
+loco,locoL_scratchsat_s162000_K1000,scratch 162k final (cross-cohort),Monocyte,12.5686,18.7801,0.669249,0.452005,0.39259,16.9171,20.8171
+loco,locoL_scratchsat_s162000_K1000,scratch 162k final (cross-cohort),T cell,17.0688,18.6943,0.913051,0.510532,0.523322,44.9667,48.4547
+loco,locoL_scratchsat_s162000_K1000,scratch 162k final (cross-cohort),B cell,12.6379,15.4945,0.815638,0.627842,0.623757,21.6707,17.2757
+loco,locoL_scratchsat_s162000_K1000,scratch 162k final (cross-cohort),NK cell,4.48981,7.26453,0.618046,0.493666,0.439817,13.5846,8.07229
+loco,locoL_scratchsat_s162000_K1000,scratch 162k final (cross-cohort),Plasma cell (L2),2.44069,3.54471,0.688544,0.891532,0.315376,1.24959,1.47858
+loco,locoL_scratchsat_s162000_K1000,scratch 162k final (cross-cohort),MEAN over 4 L1 types,null,null,null,0.521012,null,null,null
+loco,lora_n5__raw,"own-cohort pretrained LoRA (5 pairs, cross-fitted)",Monocyte,19.2603,18.7801,1.02557,0.597651,0.607988,17.2447,20.8171
+loco,lora_n5__raw,"own-cohort pretrained LoRA (5 pairs, cross-fitted)",T cell,20.4871,18.6943,1.09591,0.635655,0.609076,57.7496,48.4547
+loco,lora_n5__raw,"own-cohort pretrained LoRA (5 pairs, cross-fitted)",B cell,15.2415,15.4945,0.983671,0.77664,0.768495,17.3748,17.2757
+loco,lora_n5__raw,"own-cohort pretrained LoRA (5 pairs, cross-fitted)",NK cell,2.86171,7.26453,0.39393,0.464084,0.349683,4.20163,8.07229
+loco,lora_n5__raw,"own-cohort pretrained LoRA (5 pairs, cross-fitted)",Plasma cell (L2),0.737911,3.54471,0.208173,0.396535,0.124902,0.578862,1.47858
+loco,lora_n5__raw,"own-cohort pretrained LoRA (5 pairs, cross-fitted)",MEAN over 4 L1 types,null,null,null,0.618508,null,null,null
+loco,real_lora__raw,"own-cohort pretrained LoRA (~58 pairs, cross-fitted)",Monocyte,16.8187,18.7801,0.895559,0.672033,0.605273,18.474,20.8171
+loco,real_lora__raw,"own-cohort pretrained LoRA (~58 pairs, cross-fitted)",T cell,18.2475,18.6943,0.976102,0.706769,0.70599,55.1626,48.4547
+loco,real_lora__raw,"own-cohort pretrained LoRA (~58 pairs, cross-fitted)",B cell,15.1966,15.4945,0.980776,0.824137,0.854052,16.1756,17.2757
+loco,real_lora__raw,"own-cohort pretrained LoRA (~58 pairs, cross-fitted)",NK cell,3.9938,7.26453,0.549768,0.679397,0.721719,6.28293,8.07229
+loco,real_lora__raw,"own-cohort pretrained LoRA (~58 pairs, cross-fitted)",Plasma cell (L2),2.17479,3.54471,0.61353,0.240786,0.195338,1.08211,1.47858
+loco,real_lora__raw,"own-cohort pretrained LoRA (~58 pairs, cross-fitted)",MEAN over 4 L1 types,null,null,null,0.720584,null,null,null

paper-releases/pbisc-diffusion/updates/20261005T142225Z-rare-states-scratch-vs-pretrained/diffusion/dataset/server_1/model8/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/confound_collapse_markers.csv

@@ -0,0 +1,89 @@
+setting,parent,arm,arm_class,n_genes,median_between_donor_cv_floor_removed,median_cv_ratio_gen_over_obs,frac_genes_cv_lt_half_obs,median_within_donor_excess_dispersion,median_spearman_donor_means_vs_obs,min_n_donors
+fold0,Monocyte,OBSERVED,observed,30,0.837749,1,0,2.75278,null,17
+fold0,Monocyte,zs__raw,pretrained zero-shot,30,0.470771,0.6407,0.333333,2.83027,0.339015,17
+fold0,Monocyte,lora_n5__raw,pretrained LoRA (5 pairs),30,0.466981,0.530313,0.366667,2.83965,0.346814,17
+fold0,Monocyte,scratch_n5__raw,scratch (5 pairs),30,0.292231,0.333899,0.8,2.68103,0.219363,17
+fold0,Monocyte,real_lora__raw,pretrained LoRA (58 pairs),30,0.487888,0.623782,0.333333,3.13761,0.344363,17
+fold0,Monocyte,real_full__raw,pretrained full FT (58 pairs),30,0.449455,0.566601,0.4,2.94713,0.431373,17
+fold0,Monocyte,scratch_full__raw,"scratch (58 pairs, 100k cap)",30,0.382507,0.585894,0.333333,2.72618,0.517889,17
+fold0,Monocyte,lora_recipe__raw,pretrained LoRA + recipes (58 pairs),30,0.474512,0.498332,0.5,3.06876,0.372549,17
+fold0,Monocyte,scratch_recipe__raw,scratch + recipes (58 pairs),30,0.483099,0.708064,0.266667,3.32613,0.443956,14
+fold0,T cell,OBSERVED,observed,20,0.745895,1,0,3.62999,null,17
+fold0,T cell,zs__raw,pretrained zero-shot,20,0.666526,0.846844,0.15,3.76272,0.51348,17
+fold0,T cell,lora_n5__raw,pretrained LoRA (5 pairs),20,0.486904,0.618737,0.2,4.31793,0.79902,17
+fold0,T cell,scratch_n5__raw,scratch (5 pairs),20,0.135078,0.209019,1,4.2345,0.181373,17
+fold0,T cell,real_lora__raw,pretrained LoRA (58 pairs),20,0.682387,0.828942,0.2,7.08976,0.822304,17
+fold0,T cell,real_full__raw,pretrained full FT (58 pairs),20,0.651764,0.833558,0.25,7.17462,0.852941,17
+fold0,T cell,scratch_full__raw,"scratch (58 pairs, 100k cap)",20,0.443003,0.485269,0.5,5.88421,0.740196,17
+fold0,T cell,lora_recipe__raw,pretrained LoRA + recipes (58 pairs),20,0.671228,0.838902,0.2,6.41367,0.851716,17
+fold0,T cell,scratch_recipe__raw,scratch + recipes (58 pairs),20,0.464354,0.537365,0.5,5.23029,0.620098,17
+fold0,B cell,OBSERVED,observed,20,0.869329,1,0,5.49294,null,17
+fold0,B cell,zs__raw,pretrained zero-shot,20,0.529244,0.777994,0.4,4.62364,0.188784,17
+fold0,B cell,lora_n5__raw,pretrained LoRA (5 pairs),20,0.55706,0.607596,0.35,4.67079,0.278186,17
+fold0,B cell,scratch_n5__raw,scratch (5 pairs),20,0.219935,0.333832,0.8,5.85313,0.314951,17
+fold0,B cell,real_lora__raw,pretrained LoRA (58 pairs),20,0.601831,0.956784,0.2,6.87342,-0.0147059,17
+fold0,B cell,real_full__raw,pretrained full FT (58 pairs),20,0.534772,0.557719,0.35,7.24755,0.283088,17
+fold0,B cell,scratch_full__raw,"scratch (58 pairs, 100k cap)",20,0.394504,0.490611,0.55,6.60256,0.269608,17
+fold0,B cell,lora_recipe__raw,pretrained LoRA + recipes (58 pairs),20,0.603333,0.675102,0.25,6.10506,0.197304,17
+fold0,B cell,scratch_recipe__raw,scratch + recipes (58 pairs),20,0.446492,0.507302,0.5,8.31659,0.245098,17
+fold0,NK cell,OBSERVED,observed,24,0.523614,1,0,0.634548,null,17
+fold0,NK cell,zs__raw,pretrained zero-shot,24,0.483532,0.854716,0.208333,1.59799,0.143382,17
+fold0,NK cell,lora_n5__raw,pretrained LoRA (5 pairs),24,0.2077,0.434792,0.583333,1.1245,0.236597,17
+fold0,NK cell,scratch_n5__raw,scratch (5 pairs),24,0.100576,0.191671,0.958333,0.848542,0.00734692,17
+fold0,NK cell,real_lora__raw,pretrained LoRA (58 pairs),24,0.253018,0.502412,0.5,1.13615,0.252451,17
+fold0,NK cell,real_full__raw,pretrained full FT (58 pairs),24,0.2768,0.539631,0.416667,1.4861,0.349265,17
+fold0,NK cell,scratch_full__raw,"scratch (58 pairs, 100k cap)",24,0.194022,0.362317,0.708333,1.0733,0.345588,16
+fold0,NK cell,lora_recipe__raw,pretrained LoRA + recipes (58 pairs),24,0.2376,0.458611,0.583333,1.24787,0.259804,17
+fold0,NK cell,scratch_recipe__raw,scratch + recipes (58 pairs),24,0.202084,0.403771,0.666667,1.0456,0.305882,16
+loco,Monocyte,OBSERVED,observed,32,1.12603,1,0,2.66013,null,76
+loco,Monocyte,zs__raw,pretrained zero-shot,32,0.444234,0.577277,0.375,3.36088,0.253494,82
+loco,Monocyte,xcalL_K1000,pretrained zero-shot + cross-cohort calibration,32,0.519502,0.519988,0.4375,2.10979,0.305728,77
+loco,Monocyte,locoL_lora16sat_s26750_K1000,pretrained LoRA r16 (cross-cohort),32,0.492846,0.487279,0.53125,3.34978,0.296296,79
+loco,Monocyte,locoL_lora64sat_s59000_K1000,pretrained LoRA r64 (cross-cohort),32,0.564841,0.553855,0.28125,2.95576,0.294615,77
+loco,Monocyte,locoL_fullsat_s20000_K1000,pretrained full FT 20k (cross-cohort),32,0.503602,0.48691,0.5625,3.16453,0.243034,77
+loco,Monocyte,locoL_fullsat_s108500_K1000,pretrained full FT 108.5k saturated (cross-cohort),32,0.589046,0.492575,0.53125,2.55234,0.234423,70
+loco,Monocyte,locoL_scratchsat_s50000_K1000,scratch 50k (cross-cohort),32,0.604742,0.626745,0.34375,3.40375,0.296226,78
+loco,Monocyte,locoL_scratchsat_s100000_K1000,scratch 100k (cross-cohort),32,0.643386,0.570387,0.375,3.47431,0.308512,80
+loco,Monocyte,locoL_scratchsat_s150000_K1000,scratch 150k (cross-cohort),32,0.590276,0.581738,0.34375,3.24767,0.275346,80
+loco,Monocyte,locoL_scratchsat_s162000_K1000,scratch 162k final (cross-cohort),32,0.528723,0.589979,0.40625,3.06698,0.293845,80
+loco,Monocyte,lora_n5__raw,"own-cohort pretrained LoRA (5 pairs, cross-fitted)",32,0.889116,0.78903,0.125,3.07417,0.239237,68
+loco,Monocyte,real_lora__raw,"own-cohort pretrained LoRA (~58 pairs, cross-fitted)",32,0.624907,0.598533,0.1875,3.67658,0.331369,78
+loco,T cell,OBSERVED,observed,21,0.746737,1,0,4.09178,null,82
+loco,T cell,zs__raw,pretrained zero-shot,21,0.674623,0.85674,0.047619,3.2512,0.524605,82
+loco,T cell,xcalL_K1000,pretrained zero-shot + cross-cohort calibration,21,0.594154,0.736228,0.0952381,2.23179,0.60718,82
+loco,T cell,locoL_lora16sat_s26750_K1000,pretrained LoRA r16 (cross-cohort),21,0.587531,0.796501,0.0952381,3.82812,0.558821,82
+loco,T cell,locoL_lora64sat_s59000_K1000,pretrained LoRA r64 (cross-cohort),21,0.671181,0.883011,0.047619,3.63798,0.604728,82
+loco,T cell,locoL_fullsat_s20000_K1000,pretrained full FT 20k (cross-cohort),21,0.537268,0.749057,0.0952381,3.72927,0.602278,82
+loco,T cell,locoL_fullsat_s108500_K1000,pretrained full FT 108.5k saturated (cross-cohort),21,0.558452,0.793894,0.047619,3.24639,0.604423,82
+loco,T cell,locoL_scratchsat_s50000_K1000,scratch 50k (cross-cohort),21,0.671636,0.826204,0.047619,3.49908,0.552269,82
+loco,T cell,locoL_scratchsat_s100000_K1000,scratch 100k (cross-cohort),21,0.612239,0.758936,0,3.89528,0.550527,82
+loco,T cell,locoL_scratchsat_s150000_K1000,scratch 150k (cross-cohort),21,0.569704,0.694042,0.0952381,3.53443,0.573383,82
+loco,T cell,locoL_scratchsat_s162000_K1000,scratch 162k final (cross-cohort),21,0.561547,0.737767,0.0952381,3.73614,0.583853,82
+loco,T cell,lora_n5__raw,"own-cohort pretrained LoRA (5 pairs, cross-fitted)",21,0.758603,0.899049,0.047619,5.22264,0.520205,82
+loco,T cell,real_lora__raw,"own-cohort pretrained LoRA (~58 pairs, cross-fitted)",21,0.651563,0.850103,0.047619,5.10292,0.695715,82
+loco,B cell,OBSERVED,observed,21,1.01674,1,0,4.56312,null,82
+loco,B cell,zs__raw,pretrained zero-shot,21,0.730694,0.638747,0.238095,5.61554,0.175913,82
+loco,B cell,xcalL_K1000,pretrained zero-shot + cross-cohort calibration,21,0.931319,0.683165,0.142857,4.94875,0.203615,80
+loco,B cell,locoL_lora16sat_s26750_K1000,pretrained LoRA r16 (cross-cohort),21,0.844532,0.645469,0.190476,7.75928,0.24141,82
+loco,B cell,locoL_lora64sat_s59000_K1000,pretrained LoRA r64 (cross-cohort),21,1.06529,0.824864,0.190476,8.10297,0.237809,82
+loco,B cell,locoL_fullsat_s20000_K1000,pretrained full FT 20k (cross-cohort),21,1.05524,0.773722,0.142857,6.19991,0.228138,82
+loco,B cell,locoL_fullsat_s108500_K1000,pretrained full FT 108.5k saturated (cross-cohort),21,0.986717,0.84969,0.0952381,6.2986,0.238886,82
+loco,B cell,locoL_scratchsat_s50000_K1000,scratch 50k (cross-cohort),21,1.14271,0.750769,0.238095,8.48569,0.288132,82
+loco,B cell,locoL_scratchsat_s100000_K1000,scratch 100k (cross-cohort),21,0.964441,0.758409,0.333333,8.84602,0.247651,82
+loco,B cell,locoL_scratchsat_s150000_K1000,scratch 150k (cross-cohort),21,0.973093,0.666081,0.285714,10.4648,0.252163,82
+loco,B cell,locoL_scratchsat_s162000_K1000,scratch 162k final (cross-cohort),21,0.840098,0.681208,0.238095,8.64591,0.275563,82
+loco,B cell,lora_n5__raw,"own-cohort pretrained LoRA (5 pairs, cross-fitted)",21,1.01441,0.767579,0.047619,6.27628,0.0616136,82
+loco,B cell,real_lora__raw,"own-cohort pretrained LoRA (~58 pairs, cross-fitted)",21,1.06256,0.691948,0.238095,7.26977,0.0881031,82
+loco,NK cell,OBSERVED,observed,25,0.554711,1,0,0.684981,null,82
+loco,NK cell,zs__raw,pretrained zero-shot,25,0.470313,0.807546,0.16,1.46877,0.269341,82
+loco,NK cell,xcalL_K1000,pretrained zero-shot + cross-cohort calibration,25,0.404367,0.691789,0.2,1.7763,0.261301,79
+loco,NK cell,locoL_lora16sat_s26750_K1000,pretrained LoRA r16 (cross-cohort),25,0.214373,0.414381,0.64,1.56617,0.147637,82
+loco,NK cell,locoL_lora64sat_s59000_K1000,pretrained LoRA r64 (cross-cohort),25,0.358444,0.533157,0.36,1.761,0.242946,82
+loco,NK cell,locoL_fullsat_s20000_K1000,pretrained full FT 20k (cross-cohort),25,0.2634,0.512792,0.48,1.74234,0.256908,82
+loco,NK cell,locoL_fullsat_s108500_K1000,pretrained full FT 108.5k saturated (cross-cohort),25,0.411867,0.691865,0.2,1.26887,0.214092,81
+loco,NK cell,locoL_scratchsat_s50000_K1000,scratch 50k (cross-cohort),25,0.277024,0.524211,0.4,0.981946,0.239538,82
+loco,NK cell,locoL_scratchsat_s100000_K1000,scratch 100k (cross-cohort),25,0.280391,0.558118,0.36,1.07907,0.234333,82
+loco,NK cell,locoL_scratchsat_s150000_K1000,scratch 150k (cross-cohort),25,0.267707,0.497029,0.52,1.0292,0.272114,82
+loco,NK cell,locoL_scratchsat_s162000_K1000,scratch 162k final (cross-cohort),25,0.28822,0.516553,0.48,1.0306,0.281095,82
+loco,NK cell,lora_n5__raw,"own-cohort pretrained LoRA (5 pairs, cross-fitted)",25,0.434181,0.749051,0.08,1.22477,0.203026,74
+loco,NK cell,real_lora__raw,"own-cohort pretrained LoRA (~58 pairs, cross-fitted)",25,0.287511,0.52839,0.48,0.876394,0.261222,79