LIBERO cells: record the phase-2 init explicitly in ctrl_meta, final cards

libero_goal_frac10_terra/README.md

@@ -63,3 +63,10 @@ python -m acdc_la.downstream.vla.run_libero_eval_cont --pretrained_checkpoint ./
 ⚠ **`action_l1_1step` is oracle-conditioned** (the flow loss scores a one-step correction from a point that already
 contains the ground truth). Use it to track one run — **never to rank two arms**; on another pair in this repo it
 ranks them backwards. Rank with the suite eval.
+
+## Provenance is in the files, not just in this card
+
+`ctrl_meta.json` here records **`"init": "none"`** explicitly, because the metadata written by the training script
+does *not* otherwise say which base a checkpoint was post-trained from — a reader could not tell this arm (stock
+`openvla-7b`) from `libero_10_frac10_terra_p2fft20k` (TERRA phase-2 base) by looking at the files. The four
+directories published on 2026-11-12 record it; the older ones state it only in their cards.

libero_goal_frac10_terra/ctrl_meta.json

@@ -20,5 +20,7 @@
   "layers": 4,
   "z_tok": 32,
   "steps": 10
- }
+ },
+ "init": "none",
+ "init_note": "openvla/openvla-7b directly, NO phase 2 -- this cell predates phase 2 (LEDGER.md 2026-09-19/09-20)"
 }