fix: add top_n cutoff to ELO tournament (default top 10 generations)

ml/eval/elo_tracker.py

@@ -16,6 +16,7 @@ Reuses `run_match`, `LearnedPolicy`, `HeuristicPolicy`,
 Usage:
     python -m ml.eval.elo_tracker
     python -m ml.eval.elo_tracker --episodes 50 --output data/elo_history.json
+    python -m ml.eval.elo_tracker --top-n 10
 """
 
 import json
@@ -285,12 +286,17 @@ def run_generation_tournament(
     episodes_per_matchup: int = 50,
     playstyle: str = "midrange",
     k_factor: float = DEFAULT_K,
+    top_n: int = 10,
 ) -> EloResult:
     """
     Discover generation checkpoints and run an ELO tournament.
 
     Looks for `gen-*/best_policy.pt` under checkpoint_dir.
     Also includes heuristic and random baselines.
+
+    Args:
+        top_n: Only include the N most recent gen-* checkpoints.
+               0 means include all (unbounded). Default is 10.
     """
     policies: Dict[str, object] = {
         "heuristic": HeuristicPolicy(),
@@ -299,7 +305,20 @@ def run_generation_tournament(
 
     ckpt_path = Path(checkpoint_dir)
     if ckpt_path.exists():
-        for gen_dir in sorted(ckpt_path.glob("gen-*")):
+        gen_dirs = sorted(ckpt_path.glob("gen-*"))
+
+        # Apply top_n cutoff — keep only the latest N generation dirs
+        if top_n > 0 and len(gen_dirs) > top_n:
+            skipped = len(gen_dirs) - top_n
+            logger.info(
+                "top_n=%d: skipping %d older generation(s), using %s … %s",
+                top_n, skipped,
+                gen_dirs[-top_n].name,
+                gen_dirs[-1].name,
+            )
+            gen_dirs = gen_dirs[-top_n:]
+
+        for gen_dir in gen_dirs:
             model_path = gen_dir / "best_policy.pt"
             if model_path.exists():
                 name = gen_dir.name  # e.g., "gen-001"
@@ -363,6 +382,10 @@ def main():
         "--k-factor", type=float, default=32,
         help="ELO K-factor (default: 32)",
     )
+    parser.add_argument(
+        "--top-n", type=int, default=10,
+        help="Only include the N most recent gen checkpoints (0 = all, default: 10)",
+    )
 
     args = parser.parse_args()
 
@@ -373,6 +396,7 @@ def main():
         episodes_per_matchup=args.episodes,
         playstyle=args.playstyle,
         k_factor=args.k_factor,
+        top_n=args.top_n,
     )
 
     # Print results

routes/ml.py

@@ -1097,7 +1097,7 @@ _elo_tournament_state = _EloTournamentState()
 _elo_tournament_lock  = threading.Lock()
 
 
-def _run_elo_tournament_pipeline(episodes: int, playstyle: str):
+def _run_elo_tournament_pipeline(episodes: int, playstyle: str, top_n: int):
     """Run ELO tournament in a background thread."""
     try:
         import sys
@@ -1108,7 +1108,7 @@ def _run_elo_tournament_pipeline(episodes: int, playstyle: str):
         with _elo_tournament_lock:
             _elo_tournament_state.running = True
             _elo_tournament_state.phase = "running"
-            _elo_tournament_state.message = "Running ELO tournament..."
+            _elo_tournament_state.message = f"Running ELO tournament (top {top_n} generations)..."
             _elo_tournament_state.error = None
             _elo_tournament_state.result = None
 
@@ -1121,6 +1121,7 @@ def _run_elo_tournament_pipeline(episodes: int, playstyle: str):
             checkpoint_dir=ckpt_dir,
             episodes_per_matchup=episodes,
             playstyle=playstyle,
+            top_n=top_n,
         )
 
         # Save to ELO history
@@ -1155,29 +1156,37 @@ def _run_elo_tournament_pipeline(episodes: int, playstyle: str):
 
 @router.post("/api/ml/elo/tournament")
 async def ml_start_elo_tournament(request: FastAPIRequest):
-    """Start an ELO-rated tournament across all generation checkpoints."""
+    """Start an ELO-rated tournament across generation checkpoints.
+
+    Body params:
+        episodes (int): Games per matchup pair (default 50)
+        playstyle (str): Game playstyle (default midrange)
+        topN (int): Max number of most-recent gen checkpoints to include
+                    (default 10, 0 = unlimited)
+    """
     if _elo_tournament_state.running:
         raise HTTPException(409, "ELO tournament already in progress")
 
     body = await request.json() if await request.body() else {}
     episodes = body.get("episodes", 50)
     playstyle = body.get("playstyle", "midrange")
+    top_n = body.get("topN", 10)
 
     with _elo_tournament_lock:
         _elo_tournament_state.running = True
         _elo_tournament_state.phase = "starting"
-        _elo_tournament_state.message = "Initializing ELO tournament..."
+        _elo_tournament_state.message = f"Initializing ELO tournament (top {top_n} generations)..."
         _elo_tournament_state.result = None
         _elo_tournament_state.error = None
 
     thread = threading.Thread(
         target=_run_elo_tournament_pipeline,
-        args=(episodes, playstyle),
+        args=(episodes, playstyle, top_n),
         daemon=True,
     )
     thread.start()
 
-    return {"status": "started", "episodes": episodes}
+    return {"status": "started", "episodes": episodes, "top_n": top_n}
 
 
 @router.get("/api/ml/elo/tournament/status")