Remove unimportant MCTS adapter references

Author: jphilleo

docs/benchmark/CCERL_LEADERBOARD.md

@@ -2,7 +2,7 @@
 
 Status: current release-candidate snapshot for `CCERL-2P10-v2`.
 
-This is the current local calibration snapshot for the public Chinese Checkers engine benchmark. It is engine-relative, anchored at `random = 0`, and should not be interpreted as human Elo. Source-style open-source adapters are included directly in the all-pairs ladder, not projected from a single match. `svjayanthi-mcts-fast` is excluded from the public table because it lost every useful comparison game in screening.
+This is the current local calibration snapshot for the public Chinese Checkers engine benchmark. It is engine-relative, anchored at `random = 0`, and should not be interpreted as human Elo. Source-style open-source adapters are included directly in the all-pairs ladder, not projected from a single match.
 
 ## Fixed-Simulation Baseline Ladder
 
@@ -66,4 +66,4 @@ Interpretation: `JumpStar_60` remains the public benchmark champion for now. `Ju
 
 ## Open-Source Adapter Check
 
-Source-style adapters for representative public Chinese Checkers bots are included in the main all-pairs ladder above. The retained adapters are `marblefish-ab`, `harryz-rule`, and `zedrichu-minimax`; `svjayanthi-mcts-fast` was removed from the official ladder because it produced an uninformative all-loss profile.
+Source-style adapters for representative public Chinese Checkers bots are included in the main all-pairs ladder above. The retained adapters are `marblefish-ab`, `harryz-rule`, and `zedrichu-minimax`.

docs/benchmark/CHANGELOG_2P10_V1_TO_V2.md

@@ -27,7 +27,6 @@ during the open-source bot screening:
 | `harryz-rule` (deterministic 1-ply greedy) | 172 | 0-101-71 | 101 / 172 (**59%**) | 1470.6 | -441.7 |
 | `marblefish-ab` | 172 | 4-11-157 | 0 / 172 | 1129.8 | 985.6 |
 | `zedrichu-minimax` | 172 | 5-7-160 | 0 / 172 | 1109.5 | 1021.0 |
-| `svjayanthi-mcts-fast` | 172 | 0-5-167 | 0 / 172 | 890.6 | -585.1 |
 
 `harryz-rule` never beat the champion, yet earned an Elo only ~150 points
 below it because 59% of its games hit the ply cap with a 0.5-point draw. The
@@ -106,8 +105,6 @@ exceptions, both inflated by the v1 loophole:
 - `harryz-rule`: expected to drop from ~1471 toward the
   decisive-only estimate of ~-441 (still above strict random because some
   of its draws against weaker engines remain legitimate halves).
-- `svjayanthi-mcts-fast`: smaller correction, since most of its outcomes
-  were already decisive losses. Expected drop from ~891.
 
 Numbers will be updated on the public benchmark page after the v2 re-run
 completes.

docs/benchmark/OPEN_SOURCE_BOT_SCREENING.md

@@ -9,7 +9,6 @@ open-source projects. The most useful candidates for a first comparison were:
 
 | Project | Public description | Fit for CCERL |
 |---|---|---|
-| `SVJayanthi/ChineseCheckersMCTS` | Java Chinese Checkers player using Monte Carlo Tree Search. | Good algorithmic reference; GUI/original-code integration is brittle, so v1 uses a CCERL MCTS-style adapter. |
 | `arthur-x/MarbleFish` | Browser-playable Chinese Checkers AI using alpha-beta search and beam-style pruning. | Good algorithmic reference; v1 ports the published alpha-beta scoring/search style to CCERL. |
 | `Zedrichu/Chinese-Checkers-AI` | Python project with random, non-repeating-random, and minimax players. | Useful minimax/heuristic reference, but upstream uses `triangle_size=3`, so direct binary comparison would not be the same game. |
 | `HarryZalessky/Chinese-Checkers` | Java graph-based 121-cell, 10-piece game with a rule-based "perfect game" AI. | Good rules/heuristic reference; v1 ports the rule-path style to CCERL. |
@@ -19,7 +18,6 @@ open-source projects. The most useful candidates for a first comparison were:
 Sources:
 
 - GitHub topic page: https://github.com/topics/chinese-checkers
-- SVJayanthi MCTS project: https://github.com/SVJayanthi/ChineseCheckersMCTS
 - MarbleFish project: https://github.com/arthur-x/MarbleFish
 - Zedrichu minimax project: https://github.com/Zedrichu/Chinese-Checkers-AI
 - HarryZalessky Java project: https://github.com/HarryZalessky/Chinese-Checkers
@@ -36,7 +34,6 @@ adapters:
 | `harryz-rule` | HarryZalessky rule/path AI | deterministic forward-progress, jump, and goal-entry heuristic |
 | `zedrichu-minimax` | Zedrichu minimax heuristic players | depth-limited alpha-beta over goal occupancy and distance terms |
 | `marblefish-ab` | MarbleFish alpha-beta worker | beam-pruned alpha-beta with vertical-progress and centrality scoring |
-| `svjayanthi-mcts-fast` | SVJayanthi MCTS project | small UCT/rollout adapter over CCERL legal moves |
 
 These are not vendored upstream engines. They are source-style CCERL ports:
 the trusted CCERL referee owns legality, terminal detection, and the official
@@ -57,7 +54,7 @@ Protocol:
 - Ruleset: `CCERL-2P10-v2`
 - Position suite: all 86 rows of `official_elo_v2`
 - Pairing: color-swapped games from every row
-- Games: 172 per adapter, 688 total
+- Games: 172 per adapter, 516 total
 - Anchor model: `iter063`, 64 simulations per move
 - Runner: `tools/ccp_referee.py`
 - Direct Elo: from the adapter's perspective relative to JumpStar, draws scored as 0.5
@@ -70,14 +67,12 @@ Evaluation profile:
 | `harryz-rule` | deterministic rule heuristic |
 | `marblefish-ab` | depth 2, beam 4 |
 | `zedrichu-minimax` | depth 2, beam 4 |
-| `svjayanthi-mcts-fast` | 6 iterations, rollout depth 6 |
 
 | Opponent adapter | Games | Adapter W-D-L | Adapter score | Score CI95 | Elo vs JumpStar | Terminations |
 |---|---:|---:|---:|---:|---:|---|
 | `harryz-rule` | 172 | 0-97-75 | 0.282 | 0.245 - 0.319 | -161 | all_pieces_in_goal: 7, anti_block_goal_full: 68, max_ply: 97 |
 | `marblefish-ab` | 172 | 8-41-123 | 0.166 | 0.124 - 0.208 | -278 | all_pieces_in_goal: 107, anti_block_goal_full: 24, equal_turn_goal_draw: 4, max_ply: 37 |
 | `zedrichu-minimax` | 172 | 3-6-163 | 0.035 | 0.011 - 0.059 | -563 | all_pieces_in_goal: 157, anti_block_goal_full: 9, max_ply: 6 |
-| `svjayanthi-mcts-fast` | 172 | 0-0-172 | 0.000 | 0.000 - 0.000 | -1015 | anti_block_goal_full: 172 |
 
 The full-suite run preserves the main screening conclusion: even from the adapters' own perspective, JumpStar
 reduced to 64 simulations per move is far ahead of the first source-style
@@ -100,7 +95,6 @@ Protocol:
 
 | Opponent adapter | Games | Adapter W-D-L | Adapter score | Score CI95 approx | Elo vs Iter60 raw | Elo vs Iter60 adjusted | Terminations |
 |---|---:|---:|---:|---:|---:|---:|---|
-| `svjayanthi-mcts-fast` | 16 | 0-0-16 | 0.000 | 0.000 - 0.194 | -inf | -607 | anti_block_goal_full:16 |
 | `marblefish-ab` | 16 | 1-2-13 | 0.125 | 0.035 - 0.360 | -338 | -305 | all_pieces_in_goal:10, anti_block_goal_full:4, equal_turn_goal_draw:2 |
 | `zedrichu-minimax` | 16 | 2-0-14 | 0.125 | 0.035 - 0.360 | -338 | -305 | all_pieces_in_goal:14, anti_block_goal_full:2 |
 | `harryz-rule` | 16 | 0-10-6 | 0.312 | 0.142 - 0.556 | -137 | -128 | all_pieces_in_goal:1, anti_block_goal_full:5, max_ply:10 |
@@ -120,7 +114,7 @@ The current evidence should remain bounded:
 - This is a source-style adapter evaluation, not the official public ladder.
 - The adapters are CCERL-rule ports, not exact upstream binaries.
 - Stronger exact integrations should target `alexicanesse/ChineseCheckers`,
-  `SVJayanthi/ChineseCheckersMCTS`, and `MarbleFish` first.
+  `MarbleFish`, or other actively maintained public engines first.
 - Any exact upstream binary benchmark must document rule differences, build
   steps, commit hashes, and coordinate conversion tests before its result is
   mixed into CCERL.

docs/benchmark/PUBLIC_BENCHMARK.md

@@ -70,7 +70,7 @@ The current native/CCP all-pairs ladder is:
 experiments/public_benchmark/ladder_v2_latest_champions_s64/baseline_ladder.json
 ```
 
-It uses all 86 audited v2 positions, paired color swaps, all 171 displayed-engine pairs, 29,412 games total, 64-simulation neural search, the existing native/CCP logs plus latest-champion incremental runs, and `random = 0`. Open-source adapter rows are fitted in the same all-pairs graph as every other displayed engine; `svjayanthi-mcts-fast` is excluded because it produced no useful comparison signal.
+It uses all 86 audited v2 positions, paired color swaps, all 171 displayed-engine pairs, 29,412 games total, 64-simulation neural search, the existing native/CCP logs plus latest-champion incremental runs, and `random = 0`. Open-source adapter rows are fitted in the same all-pairs graph as every other displayed engine.
 
 | Rank | Baseline | Elo | Games | Score | vs JumpStar_60 | vs JumpStar_82 |
 |---:|---|---:|---:|---:|---:|---:|

experiments/public_benchmark/ladder_v2_latest_champions_s64/baseline_ladder.json

@@ -413,9 +413,7 @@
       }
     ]
   },
-  "excluded_labels": [
-    "svjayanthi-mcts-fast"
-  ],
+  "excluded_labels": [],
   "pairs": [
     {
       "a": "greedy",

experiments/public_benchmark/open_source_v2/elo_half.json

@@ -2,9 +2,9 @@
   "anchor": "iter063_s64",
   "anchor_elo": 0.0,
   "draw_policy": "half",
-  "games": 688,
+  "games": 516,
   "log_likelihood": -205.56415383438255,
-  "players": 5,
+  "players": 4,
   "ratings": [
     {
       "ci95": [
@@ -13,7 +13,7 @@
       ],
       "draws": 144,
       "elo": 0.0,
-      "games": 688,
+      "games": 516,
       "losses": 11,
       "player": "iter063_s64",
       "score_rate": 0.8793604651162791,
@@ -61,20 +61,6 @@
       "score_rate": 0.03488372093023256,
       "standard_error": 72.00292606070057,
       "wins": 3
-    },
-    {
-      "ci95": [
-        -4156.888254968699,
-        -269.2489590306452
-      ],
-      "draws": 0,
-      "elo": -2213.068606999672,
-      "games": 172,
-      "losses": 172,
-      "player": "svjayanthi-mcts-fast",
-      "score_rate": 0.0,
-      "standard_error": 991.7447183515442,
-      "wins": 0
     }
   ]
 }

experiments/public_benchmark/open_source_v2/results.json

@@ -62,24 +62,6 @@
         "anti_block_goal_full": 9,
         "max_ply": 6
       }
-    },
-    {
-      "anchor_losses": 0,
-      "anchor_score": 1.0,
-      "anchor_wins": 172,
-      "direct_elo_delta": null,
-      "direct_elo_delta_adjusted": 1015.127638029308,
-      "draws": 0,
-      "games": 172,
-      "log": "experiments/public_benchmark/open_source_v2/svjayanthi-mcts-fast_vs_iter063_s64.jsonl",
-      "opponent": "svjayanthi-mcts-fast",
-      "score_ci95_normal": [
-        1.0,
-        1.0
-      ],
-      "terminations": {
-        "anti_block_goal_full": 172
-      }
     }
   ],
   "elo_half": "experiments/public_benchmark/open_source_v2/elo_half.json",

tools/ccp_open_source_style_bot.py

@@ -471,16 +471,14 @@ def choose_move(args: argparse.Namespace, cells: str, player: int, rng: random.R
         return choose_alphabeta(cells, player, args.depth, marblefish_score, beam=args.beam)
     if args.style == "harryz-rule":
         return choose_rule_path(cells, player)
-    if args.style in {"svjayanthi-mcts", "svjayanthi-mcts-fast"}:
-        return choose_mcts(cells, player, rng, args.iterations, args.rollout_depth)
     raise RuntimeError(f"unknown style: {args.style}")
 
 
 def main() -> int:
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument(
         "--style",
-        choices=("zedrichu-minimax", "marblefish-ab", "harryz-rule", "svjayanthi-mcts", "svjayanthi-mcts-fast"),
+        choices=("zedrichu-minimax", "marblefish-ab", "harryz-rule"),
         required=True,
     )
     parser.add_argument("--name")