Fixed-split CE + anchor-in-I (_d512) trained on to 2,000 epochs: selected checkpoint ep1650, plan entry and README note

The campaign-2 schedule stopped this run at 1,000 epochs; it was continued from the epoch-1,000 weights with a fresh optimizer, the way the original fixed-split I-CE and CE reached 2,000. The epoch-1,000 file stays; the plan entry keeps it under 'previous'.

README.md

@@ -76,6 +76,11 @@ in `c2_docfix512_ckpt_plan.json` (same entry format, plus the file's SHA-256 and
 - Fixed split (`Pod/w9_a100_worker.py`, seed 0, the original run's epoch count): I-CE, CE and
   CE + anchor-in-I of Table 1 and the four alignment-plus-uniformity runs aimed at the anchor gallery
   (`ai25auni2_icetf`, `ai6auni2_icetf`, `ai4auni2_icetf`, `ai2auni25_icetf`); keys end in `_fixed_d512`.
+  The fixed-split CE + anchor-in-I (`ai2ce_icetf_fixed_d512`), which that schedule stopped at 1,000
+  epochs, was trained on to 2,000 on 2026-11-08 (weights of epoch 1,000, fresh optimizer, the way the
+  original fixed-split I-CE and CE reached 2,000); the rank rule over all 40 checkpoints now selects
+  `ckpt_w9_wcle_ai2ce_icetf_d512_fp_ep1650.pt`. The epoch-1,000 file stays, and the plan entry keeps it
+  under `previous`.
 
 Runs whose loss never reads the anchors (VICReg, the SimCLR-style tower, BYOL, I + SimCLR and alignment
 plus uniformity at the batch views) train the same weights on either pack, which a three-step A/B on

c2_docfix512_ckpt_plan.json

@@ -687,10 +687,17 @@
   "protocol": "fixed split (seed 0)",
   "worker": "Pod/w9_a100_worker.py",
   "sfx": "_d512",
-  "ep": 1000,
-  "ckpt": "ckpt_w9_wcle_ai2ce_icetf_d512_fp_ep1000.pt",
-  "sha256": "42ca7e5bad6f025c8b42ac7a955c90f0fe06cd850a64f52f96945ec11e4d5f5b",
-  "ranksel": 127.89447553198298
+  "ep": 1650,
+  "ckpt": "ckpt_w9_wcle_ai2ce_icetf_d512_fp_ep1650.pt",
+  "sha256": "838fd0cb8af107844fd9f0356a87825bdd46a3be76afbbf2f22c16f7e8f6ca51",
+  "ranksel": 129.90038982957003,
+  "extended": "2026-11-08: trained on from 1,000 to 2,000 epochs (weights of ep1000, fresh optimizer, AMP and RNG, as the original fixed-split I-CE and CE reached 2,000); rank rule over all 40 checkpoints",
+  "previous": {
+   "ep": 1000,
+   "ckpt": "ckpt_w9_wcle_ai2ce_icetf_d512_fp_ep1000.pt",
+   "sha256": "42ca7e5bad6f025c8b42ac7a955c90f0fe06cd850a64f52f96945ec11e4d5f5b",
+   "ranksel": 127.89447553198298
+  }
  },
  "ai4auni2_icetf_fixed_d512": {
   "arm": "ai4auni2_icetf",
@@ -736,4 +743,4 @@
   "sha256": "7955a086ee874f4eaa07432f1ee3d92b9dff76cf3d6b0f7151b05ec4c616f0d4",
   "ranksel": 130.45839742717456
  }
-}
+}

ckpt_w9_wcle_ai2ce_icetf_d512_fp_ep1650.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:838fd0cb8af107844fd9f0356a87825bdd46a3be76afbbf2f22c16f7e8f6ca51
+size 1452322