Use learned memory before seeding lucky prompts

CHANGELOG.md

@@ -28,6 +28,8 @@ Updated `I'm Feeling Lucky` filtering to learn poor adjacent token pairs instead
 
 Updated `I'm Feeling Lucky` with uncapped token, pair, and context memory so it can learn which concepts belong together and call strong missing neighbors when prompt anchors are present.
 
+Fixed `I'm Feeling Lucky` composition order so the current generation uses already-learned memory first, then seeds current prompt tokens for future runs.
+
 ## [2.1.3] - 2026-04-24
 
 ### Changed

conditioning.py

@@ -3746,22 +3746,13 @@ class FunPackVideoRefiner:
                 token_mask=bank_rated_token_mask,
             )
 
-        if analysis_prompt and analysis_prompt != bank_rated_prompt:
-            self._seed_void_token_bank(
-                global_adaptive,
-                word_groups,
-                cur_positive,
-                iter_num,
-                token_mask=active_token_mask,
-            )
-        elif bank_rated_positive is None:
-            self._seed_void_token_bank(
-                global_adaptive,
-                word_groups,
-                cur_positive,
-                iter_num,
-                token_mask=active_token_mask,
+        should_seed_current_prompt_after_lucky = (
+            bool(analysis_prompt) and
+            (
+                analysis_prompt != bank_rated_prompt or
+                bank_rated_positive is None
             )
+        )
 
         if feedback_enabled and pending is not None:
             self._apply_concept_feedback(
@@ -4070,6 +4061,14 @@ class FunPackVideoRefiner:
             token_mask=active_token_mask,
             prompt_sequences=self._lucky_prompt_sequences(prompt_histories),
         )
+        if should_seed_current_prompt_after_lucky:
+            self._seed_void_token_bank(
+                global_adaptive,
+                word_groups,
+                cur_positive,
+                iter_num,
+                token_mask=active_token_mask,
+            )
 
         # Update momentum
         momentum = 0.75 * momentum + 0.25 * (new_delta * reward)

docs/FunPackGemmaEmbeddingRefiner.md

@@ -92,7 +92,7 @@ Use it when you want to discover surprising concept-concept pairs or learn which
 
 `im_feeling_lucky` is the automated prompt-composer path. Instead of requiring the current prompt to name every action, camera move, or styling cue, it learns which tokens and concepts the user tends to prefer and applies them directly to conditioning. If the prompt contains a known anchor, Lucky can add strongly rated missing neighbours that commonly work with that anchor. If the prompt is vague, it uses global learned preferences so the result still has something intentional to look at.
 
-Lucky stores token, pair, and context memory without a fixed cap. Tokens stay available unless their embeddings are invalid; bad ratings teach the refiner to avoid specific adjacent token pairs or contexts first. With changing prompts, the previous generation's rating is applied to the previous prompt and its injected Lucky tokens, while the current prompt contributes fresh neutral tokens for Lucky to try. It can run by itself or alongside `into_the_void`.
+Lucky stores token, pair, and context memory without a fixed cap. Tokens stay available unless their embeddings are invalid; bad ratings teach the refiner to avoid specific adjacent token pairs or contexts first. With changing prompts, the previous generation's rating is applied to the previous prompt and its injected Lucky tokens, then Lucky composes the current generation from that already-learned memory. The current prompt contributes fresh neutral tokens only after current Lucky composition, so new words become material for future runs rather than steering the same run. It can run by itself or alongside `into_the_void`.
 
 ## Latent Refinement