@openpond/evals 0.12.6 → 0.12.8

package/LEARNING.md

@@ -217,6 +217,11 @@ approved-batch methods are SFT, GRPO and PPO with human admission; qualified
 automatic admission and automatic acceptance/serving are rejected until their
 execution paths are implemented.
 
+Hosted SFT continual learning schedules reviewed retraining from the pinned base
+model. An accepted SFT candidate remains review evidence; the next SFT iteration
+does not reuse its adapter or optimizer checkpoint. GRPO retains its accepted
+parent checkpoint selection and the host's continuation qualification checks.
+
 Manual triggers carry a stable `identity`; scheduled triggers carry a concrete
 `scheduledAt`. Equivalent timestamp offsets denote the same fire. A manual
 trigger can use a policy with a saved schedule without editing that policy.

package/RELEASING.md

@@ -36,3 +36,9 @@ leave the lockfile unchanged. The helper reports the selected path before commit
 A dry run requires clean, current master and prints the commands without changing
 Git or package versions. Existing trusted publishing, registry verification,
 package tags, and retry behavior are unchanged.
+
+Version `0.12.8` gives newly authored Reward support files content-addressed
+paths so training and evaluation Rewards, including different revisions of one
+Reward, can coexist in one private export. Existing published assets and checked
+Reward revisions remain immutable; edit, recheck and publish a new revision to
+use the new paths. Asset visibility and reference-integrity checks are unchanged.