{"approach":"Random-initialized 56-16-3 feed-forward ReLU network (963 float32 weights/biases), using only the current 56-number frame of the 224-number observation. Population 64, eight elites, Gaussian mutations of all or 12% of parameters with multiple scales, 20 minutes against the decoded current entry. Fresh paired random-spawn batches (8 matches initially, 16 thereafter); kills first (10000*elimination wins), effective HP difference second, board points only as a tiny tie-break. Crashes count as zero effective HP despite the simulator retaining raw HP. Every tenth generation selected on a separate 128-match paired validation set; checkpoints near 5, 10, 15 and 20 minutes then compared on another 256 paired ONNX validation matches, selecting generation 1330. All reported final records use untouched seeds and the repository run_match with default physics, 90-second limit, randomize_spawns=true, balanced sides, fresh built-in policy state per match. Scores are (eliminations+0.75*HP wins+0.5*draws)/matches, not the repository 3/2/1 points. Evolution used native inference checked against ONNX Runtime; all four final records and current_vs_starter use actual ONNX Runtime inference. Simulator source commit 10a6b40d717f54a631e833ef850fd2f1e4066798. Decoded current model SHA256 f3963181497e39089eef5862284d575d5bfc20e6e41b6377b95ee76c436e25b4; its separate JSON SHA metadata is inconsistent. Observed on the final current-entry benchmark: it camps near the ceiling at low speed (time-weighted mean altitude 556.3 m, speed 44.4 m/s, 82.7% of sampled frames above 550 m) and fires frequently (40.34 shots/match versus 7.457 for the current model). This is a specialized counter-policy: its strong target matchup does not carry over to Starter, the built-ins, or the fresh exploiter. It is useful as a different sparring opponent, and its broad benchmark record does not justify replacing the current upload.","blocked":null,"current_vs_starter":0.9923333333333333,"records":{"starter":{"matches":3000,"score":0.20925},"builtins":{"matches":800,"score":0.003125},"current":{"matches":1000,"wins":603,"losses":395,"draws":2,"score":0.60375},"exploiter":{"matches":500,"score":0.0,"how":"Two searches ran in parallel with a 600-second budget, two CPU threads each, only against the frozen final model: a small-network population half initialized from the final policy plus noise and half random, and a Starter-output-head population. Population 64, eight elites, multiscale Gaussian mutations, kills first and effective HP second; fresh paired spawn batches, separate 128-match native validation every ten generations. The head search stopped early enough to keep its expensive generations within the window. Five- and ten-minute checkpoints were compared on 256 other paired ONNX matches. Strongest selected was a 56-16-3 ReLU network with all 963 parameters evolved, checkpoint net-10min, seed 8653, generation 420; its ONNX validation score was 1.0 from the exploiter side. The reported 500-match score is from the final model side on untouched seeds."}},"entry":"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","entry_sha256":"18a05bf4a5c1321e263367c28c715c4248d7785b1f36def5000d84780a5c85ed","entry_bytes":4677,"ops":["Gather","Gemm","Relu"],"tried":["Harness: decoded current vs Starter, fresh seeds 400000000..400001499, both sides: 2977 kills, 0 HP wins, 0 draws, 23 losses /3000; score 0.9923333333333333. Prior reported score was 0.9933333333333333.","Early kill-minus-loss objective: passive circling, 0 kills, 190 draws and 10 losses /200 untouched matches, score 0.475. Rejected; restart from random initialization with kills primary and corrected crash HP.","Native-proxy kills-first run, roughly two-minute snapshot on 1000 additional untouched matches: 320 kills, 1 HP wins, 1 draws, 678 losses; score 0.32125.","Main kills-first run: 1200.244986517 seconds, 3010 generations, 3,066,880 native training matches plus 38,528 separate native validation matches. Validation was used for selection; final benchmark seeds were never used for training or checkpoint selection.","Actual ONNX checkpoint validation, 256 paired matches on seeds 600000000..600000127: 5min: kills=156, HP wins=0, draws=1, losses=99, score=0.611328125; 10min: kills=161, HP wins=0, draws=1, losses=94, score=0.630859375; 15min: kills=162, HP wins=0, draws=2, losses=92, score=0.63671875; 20min: kills=162, HP wins=0, draws=2, losses=92, score=0.63671875. Twenty-minute weights were byte-identical to fifteen-minute weights, so their validation record was reused.","Generalization warning found before freezing: five-minute checkpoint vs Starter on 200 untouched ONNX matches: 9 kills, 2 HP wins, 5 draws, 184 losses; score 0.065. Training remained against the current model as assigned.","Checks: all 65 repository simulator tests passed. Current and Starter native inference matched ONNX Runtime within 1.8e-7 and 3.4e-6 respectively on 10,000 random observations, with no shoot-sign differences. Small floating-point differences can alter match trajectories, so all final records used actual ONNX Runtime. Final ONNX passed full ONNX checking, permitted-op subset checking, and 10,000 finite-output checks; maximum NumPy forward-pass discrepancy 4.57763671875e-05 with zero shoot-sign differences.","Final built-ins, each 200 fresh paired ONNX matches: dogfighter: 0 kills, 0 HP wins, 0 draws, 200 losses, score 0.0; chaser: 2 kills, 0 HP wins, 0 draws, 198 losses, score 0.01; ace: 0 kills, 0 HP wins, 0 draws, 200 losses, score 0.0; brawler: 0 kills, 0 HP wins, 1 draws, 199 losses, score 0.0025. Combined score 0.003125.","Ten-minute counter-training, separate 256-match ONNX selection from challenger side: net-5min: kills=250, draws=1, score=0.990234375; net-10min: kills=256, draws=0, score=1.0; head-5min: kills=244, draws=0, score=0.9853515625; head-10min: kills=255, draws=0, score=0.99609375. Selected net-10min.","BEST targeted sparring network: checkpoint 15min, generation 1330. On fresh ONNX benchmarks: Starter 45 kills/335 HP wins/663 draws/1957 losses in 3000, score 0.20925; built-ins score 0.003125 over 800; current 602 kills/1 HP wins/2 draws/395 losses in 1000, score 0.60375; final-model side versus selected exploiter 0 kills/0 HP wins/0 draws/500 losses in 500, score 0.0."],"next":"Retain the current upload. Add this compact ceiling-camping neural policy to a shared opponent league alongside Starter, all four built-ins, the current rule and independently trained neural policies. Train and select on the whole pool, with separate paired random-spawn validation and fresh ten-minute counter-training; preserve the current near-perfect Starter performance while removing this specific neural blind spot. The reported current matchup alone is insufficient to select a replacement."}