Archive 3 checkpoints (1-3)

paper-releases/pbisc-diffusion/updates/20261003T103040Z-crosscohort-k1000-saturation/diffusion/metadata/NOTE.md

@@ -0,0 +1,5 @@
+# PBISC-diffusion update 20261003T103040Z-crosscohort-k1000-saturation
+
+Cross-cohort (LOCO) benchmark, K = 1000 complete (2026-10-03; plans/paired_crosscohort_20261002): Liu-default training pools, step-saturation protocol (20k blocks, stop below 1 % block improvement) for LoRA r16, LoRA r64 and full fine-tuning on all five held-out cohorts, every milestone sampled and scored, plus the zero-pair cross-cohort calibration control; LoRA checkpoint integrity audit (180 runs, all pass); server_2 memory-guard stop of two full runs (rerun on server_1, partial logs kept). Score tables, aggregates and run receipts; generated NPZs as locations (scored ones moved to server_2 with sha checks); the K = 2000 units still queued are registered as planned. Not uploaded: fine-tuned checkpoints (sha256 in RUN.json).
+
+Payload states are recorded as supplied. Inference locations are pointers, not archived generated outputs.