Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261003T103040Z-crosscohort-k1000-saturation/diffusion/dataset/server_1/model8/plans/paired_crosscohort_20261002/README.md

@@ -0,0 +1,191 @@
+# Cross-cohort fine-tuning benchmark (leave-one-cohort-out), 2026-10-02
+
+**Question.** Can the final model (f01 6k · 1× · E16), fine-tuned on paired bulk–sc data of *other* cohorts, generate
+cells from the real bulk of a cohort it has never seen, with no pair from that cohort? Requested by the user on
+2026-10-02; fine-tuning modes LoRA r16, LoRA r64 and full; recipe size K = 1000 and 2000. The metrics are the
+within-cohort benchmark's (`plans/paired_finetune_intradataset_20260927`).
+
+## Design
+
+**Cohorts** (the within-cohort benchmark's five; same QC, identity checks, usable rows and cells):
+
+| cohort | bulk | usable pairs | donors | role |
+|---|---|---:|---:|---|
+| ImMicro | PBMC | 107 | 107 | train and test |
+| Ye 2026 sepsis | PBMC | 82 | 82 | train and test |
+| Imperial ESKD COVID | PBMC | 157 | 57 | train and test |
+| Giroux 2022 | PBMC | 24 | 9 | train and test |
+| Ji 2025 ChILI | whole blood | 33 | 20 | test only |
+
+**Splits.** Each split trains one model:
+- **`ho_<X>`, X a PBMC cohort:** the model trains on every usable pair of the other three PBMC cohorts. It is tested on every usable pair of X.
+- **`ho_ji2025_chili`:** the model trains on all four PBMC cohorts. This is also the candidate "pan-cohort" model to release.
+
+Ji is not used for training. Its bulk is whole blood while its single cells are PBMC, so its pairs teach a different mapping than every other cohort's.
+
+**Leakage.** No sample, cell, condition or calibration value of the held-out cohort is read when a split's training inputs are built. Both `prepare_loco.py` and the trainer assert this.
+
+**Training** (`tools/ft_sdd_multi.py`). The loss, optimiser and EMA are the same as `ft_sdd.py`.
+- **Learning rate:** LoRA 5e-4 (r16 with α 16, r64 with α 64); full fine-tune 5e-5.
+- **Steps and early stopping:** evaluation every 250 steps, patience 8, cap 20,000 steps.
+- **Batches are cohort-balanced.** Each cell picks a training cohort uniformly, then a pair, then a cell. Otherwise Imperial (146 training pairs) would dominate and Giroux (24) would barely count.
+- **Inner validation:** 10 % of each training cohort's donors per group, always the same donors (ImMicro 12, sepsis 9, Imperial 5 donors / 11 rows, Giroux 0). The early-stopping loss is the mean over cohorts, so it is cohort-balanced too.
+
+**Conditions.**
+- **Fixed-depth real bulk at recipe size K.** Counts = CPM / 1e6 × K × 3,869.6, and `recipe_k` = K.
+- **Same K for training and testing.** K = 1000 is the protocol so far; its arrays equal the within-cohort benchmark's byte for byte.
+- **K = 2000 tests a different input depth.** It changes the gene features log1p(raw · K / total); the depth-per-cell scalar is unchanged.
+
+**Arms scored per held-out cohort.** All arms score the cohort's usable samples against the same frozen reference.
+
+| arm | target pairs used | what it is |
+|---|---|---|
+| `zs_K1000`, `zs_K2000` | 0 | zero-shot, real bulk |
+| `xcal_K1000`, `xcal_K2000` | 0 | zero-shot + **cross-cohort** gene-wise calibration (α from the training cohorts' pairs, cohort-balanced). This is the control for "LOCO only learns a per-gene correction" |
+| `loco_{lora16,lora64,full}_K{1000,2000}` | 0 | fine-tuned on the other cohorts |
+| within-cohort `zs__cal`, `real_lora__raw`, `real_full__raw` | all (5-fold) | reference ceilings from the within-cohort benchmark (read, not rerun) |
+
+**Metrics** (`score_ft.py`, one call per arm):
+- composition L1 (scANVI, frozen f01common);
+- precision / coverage (and depth-matched);
+- pseudobulk r;
+- disease logFC r with the bulk-broadcast baseline and partial r;
+- top-50 DEG overlap.
+
+Giroux's disease metrics exclude the 196 sex-linked genes, as in the deck. Composition gets a paired Wilcoxon over samples against `zs_K1000` and against the within-cohort LoRA.
+
+## Running (both servers)
+
+- `queues/loco_20261002.txt` has 45 units in priority order:
+  1. K = 1000 LoRA r16 for all 5 splits;
+  2. its cross-cohort calibration control;
+  3. LoRA r64 and full at K = 1000;
+  4. K = 2000 zero-shot, calibration and the three modes.
+- **`tools/run_loco_queue.sh`** runs the same queue on both hosts. A host claims a unit (atomic `mkdir` in `claims/` on server_1) only after its GPU gate passes, so no unit runs twice.
+  - server_1 walks the queue forward and server_2 backward.
+  - server_2 copies finished runs and outputs back to the same absolute paths on server_1.
+  - A failed unit is marked FAILED and not retried; two failures in a row stop the queue.
+  - Every GPU step goes through `ml-run`. On server_1 it alternates on the shared lock with the scSemiProfiler CV queue.
+- **`tools/score_pending.sh`** (`loco-score-s1`, server_1, CPU) scores each arm as soon as it exists. A failed call leaves `<out>.FAILED` and is not retried.
+- **`tools/aggregate_loco.py`** writes `scores/AGGREGATE.md` and `scores/aggregate.csv`.
+
+Units: `loco-queue-s1` and `loco-score-s1` on server_1, `loco-queue-s2` on server_2. Started 2026-10-02 20:35 KST. server_2 waits at its gate while another user's vLLM occupies its GPU (96 %, 91 GB).
+
+**GPU time (estimate):**
+- A training run takes ~16 min at the 20k cap (LoRA, ~21 steps/s) and somewhat more for full fine-tuning.
+- Sampling one split's test set takes 2–11 min.
+- Total ≈ 12 GPU h: 30 fine-tunes ≈ 8 h, 45 sampling sets ≈ 4 h.
+- On server_1 alone, sharing the lock with the scSemiProfiler CV queue, expect about 1–1.5 days.
+
+## Update 2026-10-02 21:45: Liu 2021 is in every training pool (user decision)
+
+**What changed.** The user asked to add Liu 2021 (GSE161918) as a training-only cohort, to stop all training, and to make Liu the default. The first queue (`queues/loco_20261002.txt`, splits `ho_*`) was stopped at 21:30. Its 3 finished units (K = 1000, LoRA r16, held out Giroux / Ji / sepsis) are kept as a no-Liu reference. The ImMicro unit had trained for a few seconds; it is marked `STOPPED` and has no checkpoint.
+
+**Liu 2021.**
+- Built by `processing/liu2021_v1_20261002/` → `plans/paired_finetune_intradataset_20260927/data/liu2021/`, using the same QC bundle, gene axis and identity checks as the other cohorts. See its `QC_REPORT.md`.
+- **Bulk is PBMC:** unsorted PBMC, TruSeq, htseq.
+- **Single cells:** unsorted fraction only; hashtag and demuxlet singlets.
+- **Size:** 65 pairs, 44 subjects; 58 usable rows from 41 subjects.
+- Never a test cohort: its single cells are in the pretraining corpus.
+
+**Training exclusion** (`tools/prepare_loco_extra.py`, `EXCLUDE`, read from the processing run's `hto_negative_sensitivity.csv`).
+- Rule: 14 rows lost > 30 % of their unsorted cells to hashtag-negative calls (weak pools B2 HTO3 and B3 HTO5); they are not trained on.
+- Why: negatives carry about 20 points fewer classical monocytes, so these rows' retained cells over-represent monocytes against the all-cell bulk.
+- Liu's pool is 44 rows from 37 donors: 40 training rows plus 4 inner-validation donors.
+
+**Splits `hoL_<X>`.** Each is `ho_<X>`'s pool plus Liu. Test samples and test files are unchanged. The cross-cohort calibration is refitted on the enlarged pool.
+
+| split | training rows | inner-val rows | test |
+|---|---:|---:|---:|
+| hoL_immicro | 283 | 24 | 107 |
+| hoL_ye2026_sepsis | 305 | 27 | 82 |
+| hoL_imperial_eskd | 232 | 25 | 157 |
+| hoL_giroux2022 | 354 | 36 | 24 |
+| hoL_ji2025_chili | 378 | 36 | 33 |
+
+**Queue.** `queues/loco_liu_20261002.txt` has 45 units in the same order as before. It was launched at 21:45 on server_1 (forward) and server_2 (backward, waiting at its gate). Arms are scored as `locoL_<tag>_K<K>` and `xcalL_K<K>`.
+
+**No-Liu reference so far** (K = 1000, LoRA r16; `scores/AGGREGATE.md`):
+
+| held out | coverage, zs → LOCO (within-cohort LoRA) | comp. L1, zs → LOCO (within) | logFC r, zs → LOCO (within) |
+|---|---|---|---|
+| Giroux | 0.29 → 0.81 (0.93) | 0.337 → 0.347, p 0.83 (0.258) | 0.34 → 0.30 (0.46) |
+| Sepsis | 0.07 → 0.29 (0.83) | 0.509 → 0.538, p 0.18 (0.362) | −0.02 → 0.09 (0.53) |
+| Ji | 0.06 → 0.31 (0.86) | 0.725 → 0.494, p 6e-4 (0.504) | 0.07 → 0.09 (0.48) |
+
+Pairs from other cohorts recover part of the cell realism. In Ji they also fix composition, as much as Ji's own pairs. They barely move the disease signal, which so far needs the cohort's own pairs.
+
+## Update 2026-10-03 02:00: step-saturation protocol (user decision)
+
+**Why.** Every 20k-capped run had stopped at the cap while its inner-validation loss was still falling. The user asked to keep training until it saturates: run to 40k, then 60k if it still improves, and so on. LoRA r16 and r64 come first, to check saturation before full fine-tuning.
+
+**Protocol** (`ft_sdd_multi.py --block 20000 --max-steps 200000 --saturation-frac 0.01`).
+- Training runs in 20k-step blocks.
+- At every block end, the best-so-far EMA state is saved as `checkpoint_ft_s<step>.pt` (LoRA merged).
+- The run stops when a block lowers the best inner-validation loss by less than 1 % of the total improvement over the pretrained model. Patience early stopping still applies inside a block.
+- **Each milestone is sampled and scored** as `locoL_<tag>_s<step>_K<K>`, so test metrics are shown as a function of training length.
+- **The stopping decision uses only inner-validation loss** (training cohorts), never test scores.
+- Tags are `lora16sat`, `lora64sat` and `fullsat`. The milestone checkpoint equals the final state (checked; CPU smokes for LoRA r64 and full).
+
+**Queue.** `queues/loco_liu_sat_20261003.txt` has 40 units.
+- Order: K = 1000 r16, then r64, then full; then K = 2000 (zero-shot, calibration, r16, r64, full).
+- Both hosts walk it forward.
+- Launched 02:01. `loco_liu_20261002.txt` was stopped at 01:56. Its finished 20k-cap units are kept: r16 for all five splits, r64 for Giroux and Ji, and the K = 1000 calibration arms. The sepsis r64 run was interrupted at step 15,750 and is marked `STOPPED`.
+
+**scSemiProfiler.** The CV / budget queue stays paused (`loco-pause-scsemi`) until both LOCO queues finish.
+
+## Step saturation, LoRA r16 and r64 at K = 1000 (finished 2026-10-03 10:13)
+
+| held out | r16 stop (best step) | r64 stop (best step) | L1 r16 / r64 / xcal | coverage r16 / r64 / xcal | logFC r r16 / r64 / xcal |
+|---|---|---|---|---|---|
+| ImMicro | 28,500 (26,500) | 39,000 (37,000) | 0.175 / 0.199 / 0.172 | 0.45 / 0.38 / 0.57 | – |
+| Sepsis | 26,750 (24,750) | 59,000 (57,000) | 0.491 / 0.570 / 0.444 | 0.34 / 0.29 / 0.46 | 0.10 / 0.23 / 0.10 |
+| Imperial | 40,000 (39,750) | 69,250 (67,250) | 0.635 / 0.594 / 0.651 | 0.68 / 0.72 / 0.64 | 0.49 / 0.58 / 0.69 |
+| Giroux | 35,750 (33,750) | 69,250 (67,250) | 0.431 / 0.410 / 0.202 | 0.82 / 0.81 / 0.89 | 0.30 / 0.28 / 0.24 |
+| Ji | 40,000 (38,250) | 51,750 (49,750) | 0.740 / 0.620 / 0.565 | 0.34 / 0.40 / 0.07 | 0.05 / 0.16 / 0.13 |
+
+**Saturation.**
+- Every r16 run stopped by 40k: after 20k, the best inner-validation loss improved by less than 1 % of the total.
+- r64 kept improving 2-3 % per block up to 40-60k and stopped at 39-69k.
+
+**Test metrics vs training length.**
+- **r16:** within ±0.02 of its 20k value.
+- **r64:** improved with training in Giroux composition (0.49 → 0.41) and in Imperial on every metric. Coverage fell in ImMicro.
+- **Disease signal:** in r16 it tends to fall with training (Ji 0.14 → 0.05).
+
+**r64 vs r16.**
+- Better in Imperial and Ji.
+- Worse in ImMicro, and in sepsis composition and coverage, although sepsis disease signal is higher.
+- Similar in Giroux.
+
+**Against the zero-pair calibration control and the cohort's own pairs.**
+- Neither rank beats the calibration on composition, except Imperial r64.
+- None comes near the cohort's own-pair LoRA.
+
+Full fine-tuning (same protocol) started at 10:13.
+
+## K = 1000 complete (2026-10-03 19:28): all three modes at step saturation
+
+Source: `scores/k1000_saturated_summary.csv`; each model's last milestone. Steps: LoRA r16 26-40k, r64 39-69k, full 65-120k.
+
+| held out | best on composition (L1) | best on coverage | best on disease logFC r |
+|---|---|---|---|
+| ImMicro | calibration 0.172 (r16 0.175, r64 0.199, full 0.213) | calibration 0.57 (r16 0.45) | – |
+| Sepsis | calibration 0.444 (r16 0.491, r64 0.570, full 0.585) | calibration 0.46 (r16/full 0.34) | r64 0.23 (others 0.07-0.10) |
+| Imperial | full 0.580 / r64 0.594 (calibration 0.651) | r64 0.72 (calibration 0.64, full 0.61) | calibration 0.69 (r64 0.58, full 0.46) |
+| Giroux | calibration 0.202 (r64 0.410, full 0.465) | calibration 0.89 (r16 0.82) | r16 0.30 (calibration 0.24, full 0.20) |
+| Ji (whole blood) | calibration 0.565 (r64 0.620, full 1.030) | r64 0.40 (calibration 0.07, full 0.08) | r64 0.16 (calibration 0.13) |
+
+**Findings.**
+- **Full fine-tuning never wins.** Its test metrics fall as training continues, even though the training cohorts' inner-validation loss keeps falling. Examples: Ji L1 0.84 → 1.03, ImMicro coverage 0.39 → 0.35, sepsis partial r 0.05 → 0.00. Full fine-tuning on the cohort's own pairs is among the best arms (ImMicro L1 0.157, sepsis logFC r 0.53), so its extra capacity helps only when it trains on the target cohort.
+- **LoRA rank.** r64 beats r16 in Imperial and Ji, ties in Giroux, and loses in ImMicro and on sepsis composition.
+- **Calibration control.** The zero-pair cross-cohort calibration is best or tied on composition in 4/5 cohorts and on coverage in 3/5.
+- **Own pairs.** Nothing that uses zero target pairs comes near the cohort's own-pair LoRA on any metric (L1 0.16-0.50, coverage 0.78-0.86, logFC r 0.46-0.82).
+
+## Files
+
+- `tools/prepare_loco.py` → `data/<cohort>/`, `splits/`, `PREPARE.json` (CPU; run 2026-10-02 20:31, all checks true)
+- `tools/ft_sdd_multi.py`: trainer (CPU smoke and 20-step GPU smoke passed)
+- `runs/<split>/K<K>/<tag>/` (RUN.json, history.json with per-cohort inner-val loss, checkpoint)
+- `generated/<cohort>/{zs_K2000.npz, <split>/xcal_K<K>.npz, <split>/K<K>_<tag>.npz}` + receipts
+- `scores/<cohort>/<arm>/`, `scores_nosex/giroux2022/<arm>/`

paper-releases/pbisc-diffusion/updates/20261003T103040Z-crosscohort-k1000-saturation/diffusion/dataset/server_1/model8/plans/paired_crosscohort_20261002/queues/loco_liu_sat_20261003.txt

@@ -0,0 +1,45 @@
+# Cross-cohort LOCO with Liu (splits hoL_*), step-saturation protocol (user decision 2026-10-03 01:55): every
+# fine-tune runs in 20k-step blocks and continues while a block lowers the best inner-val loss by >= 1 % of the total
+# improvement (cap 200k); every block-end checkpoint is sampled and scored. Order: LoRA r16 and r64 at K = 1000 first
+# (check saturation), then full fine-tuning; then K = 2000 (zero-shot, calibration, three modes).
+# Replaces queues/loco_liu_20261002.txt (20k cap; its finished units stay as the 20k-cap results).
+loco|hoL_giroux2022|1000|lora16sat|lora|5e-4|--lora-rank|16|--lora-alpha|16|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_ji2025_chili|1000|lora16sat|lora|5e-4|--lora-rank|16|--lora-alpha|16|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_ye2026_sepsis|1000|lora16sat|lora|5e-4|--lora-rank|16|--lora-alpha|16|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_immicro|1000|lora16sat|lora|5e-4|--lora-rank|16|--lora-alpha|16|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_imperial_eskd|1000|lora16sat|lora|5e-4|--lora-rank|16|--lora-alpha|16|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_giroux2022|1000|lora64sat|lora|5e-4|--lora-rank|64|--lora-alpha|64|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_ji2025_chili|1000|lora64sat|lora|5e-4|--lora-rank|64|--lora-alpha|64|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_ye2026_sepsis|1000|lora64sat|lora|5e-4|--lora-rank|64|--lora-alpha|64|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_immicro|1000|lora64sat|lora|5e-4|--lora-rank|64|--lora-alpha|64|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_imperial_eskd|1000|lora64sat|lora|5e-4|--lora-rank|64|--lora-alpha|64|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_giroux2022|1000|fullsat|full|5e-5|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_ji2025_chili|1000|fullsat|full|5e-5|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_ye2026_sepsis|1000|fullsat|full|5e-5|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_immicro|1000|fullsat|full|5e-5|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_imperial_eskd|1000|fullsat|full|5e-5|--block|20000|--max-steps|200000|--saturation-frac|0.01
+zs|giroux2022|2000
+zs|ji2025_chili|2000
+zs|ye2026_sepsis|2000
+zs|immicro|2000
+zs|imperial_eskd|2000
+xcal|hoL_giroux2022|2000
+xcal|hoL_ji2025_chili|2000
+xcal|hoL_ye2026_sepsis|2000
+xcal|hoL_immicro|2000
+xcal|hoL_imperial_eskd|2000
+loco|hoL_giroux2022|2000|lora16sat|lora|5e-4|--lora-rank|16|--lora-alpha|16|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_ji2025_chili|2000|lora16sat|lora|5e-4|--lora-rank|16|--lora-alpha|16|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_ye2026_sepsis|2000|lora16sat|lora|5e-4|--lora-rank|16|--lora-alpha|16|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_immicro|2000|lora16sat|lora|5e-4|--lora-rank|16|--lora-alpha|16|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_imperial_eskd|2000|lora16sat|lora|5e-4|--lora-rank|16|--lora-alpha|16|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_giroux2022|2000|lora64sat|lora|5e-4|--lora-rank|64|--lora-alpha|64|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_ji2025_chili|2000|lora64sat|lora|5e-4|--lora-rank|64|--lora-alpha|64|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_ye2026_sepsis|2000|lora64sat|lora|5e-4|--lora-rank|64|--lora-alpha|64|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_immicro|2000|lora64sat|lora|5e-4|--lora-rank|64|--lora-alpha|64|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_imperial_eskd|2000|lora64sat|lora|5e-4|--lora-rank|64|--lora-alpha|64|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_giroux2022|2000|fullsat|full|5e-5|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_ji2025_chili|2000|fullsat|full|5e-5|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_ye2026_sepsis|2000|fullsat|full|5e-5|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_immicro|2000|fullsat|full|5e-5|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_imperial_eskd|2000|fullsat|full|5e-5|--block|20000|--max-steps|200000|--saturation-frac|0.01

paper-releases/pbisc-diffusion/updates/20261003T103040Z-crosscohort-k1000-saturation/diffusion/dataset/server_1/model8/plans/paired_crosscohort_20261002/queues/loco_liu_sat_rerun_full_20261003.txt

@@ -0,0 +1,5 @@
+# Rerun on server_1 of two K = 1000 full fine-tunes started on server_2 (2026-10-03 11:10-12:18): ji was stopped by the
+# server_2 memory guard at 11:16 (another user's vLLM grew); sepsis was stopped by the coordinator at 12:18 for the same risk.
+# Their partial logs are in runs/<split>/K1000/fullsat_server2_attempt_20261003/.
+loco|hoL_ji2025_chili|1000|fullsat|full|5e-5|--block|20000|--max-steps|200000|--saturation-frac|0.01
+loco|hoL_ye2026_sepsis|1000|fullsat|full|5e-5|--block|20000|--max-steps|200000|--saturation-frac|0.01

paper-releases/pbisc-diffusion/updates/20261003T103040Z-crosscohort-k1000-saturation/diffusion/dataset/server_1/model8/plans/paired_crosscohort_20261002/tools/wait_new_scores.sh

@@ -0,0 +1,12 @@
+#!/usr/bin/env bash
+# Exit (printing them) as soon as at least one arm gets a summary.json that was not there at start; also exit when
+# both LOCO queues and the scorer are inactive. Used to report each new score to the user right away.
+P=/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_crosscohort_20261002
+snap() { ls -d "$P"/scores/*/*/summary.json "$P"/scores_nosex/*/*/summary.json 2>/dev/null | sort; }
+before=$(snap)
+while true; do
+  sleep 20
+  new=$(comm -13 <(echo "$before") <(snap))
+  if [[ -n $new ]]; then sleep 90; comm -13 <(echo "$before") <(snap) | sed "s|$P/||"; exit 0; fi
+  systemctl --user is-active --quiet loco-queue-s1 || systemctl --user is-active --quiet loco-score-s1 || { echo "queues and scorer inactive"; exit 0; }
+done