Add toggle to disable match_letters assignment pass (default off) for isolation

script.py

@@ -40,6 +40,11 @@ SYMBOLIC_ONLY = False
 # future submission cycle.
 LEAN_MODE = os.environ.get("IOL_LEAN", "0") == "1"
 
+# MATCH_ASSIGNMENT (default False = disabled). When off, match_letters puzzles
+# are answered by the normal free-form LLM pass. Off for now to isolate whether
+# the assignment pass helped or hurt; flip on with IOL_MATCH_ASSIGN=1.
+MATCH_ASSIGNMENT = os.environ.get("IOL_MATCH_ASSIGN", "0") == "1"
+
 # The eval sandbox has no internet; only go offline when loading local
 # weights so Colab testing with a Hub MODEL_ID still downloads normally.
 if MODEL_ID == "." or Path(MODEL_ID).exists():
@@ -102,7 +107,8 @@ def main(test_path: str = TEST_CSV, out_path: str = OUT_CSV) -> None:
         results = run_pipeline(rows, client, budget,
                                llm_batch=LLM_BATCH, max_new_tokens=MAX_NEW_TOKENS,
                                checkpoint=lambda rs: write_submission(rs, out_path),
-                               lean=LEAN_MODE)
+                               lean=LEAN_MODE,
+                               use_match_assignment=MATCH_ASSIGNMENT)
         write_submission(results, out_path)
         print(f"wrote {out_path}: {len(results)} rows in {budget.elapsed():.1f}s",
               flush=True)

solver/pipeline.py

@@ -142,7 +142,8 @@ def run_pipeline(rows: Sequence[dict], client: Optional[LLMClient] = None,
                  conf_keep: float = CONF_KEEP, llm_batch: int = LLM_BATCH,
                  max_new_tokens: Optional[int] = None,
                  checkpoint: Optional[Callable[[List["PuzzleResult"]], None]] = None,
-                 lean: bool = False
+                 lean: bool = False,
+                 use_match_assignment: bool = True
                  ) -> List[PuzzleResult]:
     """`checkpoint`, when given, is called with the (complete, valid) results
     after the symbolic pass and after every LLM batch — so a crash at ANY
@@ -205,7 +206,11 @@ def run_pipeline(rows: Sequence[dict], client: Optional[LLMClient] = None,
     # (scores ~0). Solve it as an assignment from the model's own distribution
     # instead. Solved puzzles get high confidence so the free-form LLM pass
     # skips them; a declined puzzle falls through to that pass unchanged.
-    if getattr(client, "can_score", False):
+    # Gated by use_match_assignment: when off, match_letters puzzles go through
+    # the normal free-form LLM pass (used to isolate this pass's effect).
+    if not use_match_assignment:
+        log("match_letters assignment pass disabled; using free-form LLM path")
+    if use_match_assignment and getattr(client, "can_score", False):
         n_assigned = 0
         for i, p in enumerate(puzzles):
             if p is None or p.task_type != "match_letters" or budget.exhausted():