One prompt. One Pong player.

Design a fixed neural network controller that wins as many matches as possible in the Pong variant below. Return the actual numeric weights and biases in one response.

Architecture: 6 → 12 → 4 → 3 → 3. Apply ReLU after every hidden layer; the three output scores are linear. Each dense layer computes output = W * input + b. Use only the listed layers. Return JSON {"layers":[{"weights":[[...]],"bias":[...]}, ...]}, ordered from input to output. Each weights matrix has one row per output neuron and one column per input to that layer. Include all coefficients and biases as finite numbers. Do not return code, formulas in place of numbers, or prose.

Rules:
- Coordinates are x along the length of a 360 by 160 court, and y across it. Each policy sees itself at the left end. The ball contacts the left paddle plane at x=15 and the right paddle plane at x=345. Paddle centers are limited to y=15..145. Each paddle has a half-width of 15 along y.
- The six inputs, in order, are (ballX-180)/180, (ballY-80)/80, ballVx/6, ballVy/6, (ownPaddleY-80)/80, (opponentPaddleY-80)/80. For the right player, reflect ballX about x=180 and negate ballVx before this normalization. Y and vy retain their signs. Negative observed vx means the ball is coming toward you.
- Both policies choose simultaneously from the same pre-tick state. Outputs in order mean move y by -2, 0, or +2. The largest score wins, with ties resolved in favor of the earliest output. Move both paddles and clamp their positions first, then advance the ball by (vx,vy).
- If the updated ball y is below 4, reflect it to 8-y. If it is above 156, reflect it to 312-y. In either case negate vy. Then check paddle contact when the ball has reached or crossed its destination plane.
- At contact let offset be updated ballY minus updated receiving paddleY. A miss (absolute offset greater than 15) ends the rally and awards one point to the other player. On a hit, snap x to the contact plane. Reverse vx with new magnitude s=min(13, abs(vx)+1). Set vy=round(9*s*offset/(15*13)), where half ties round toward positive infinity.
- Each rally starts with the ball at (180,80) and both paddles at y=80. The serve has abs(vx)=5 and abs(vy) between 0.89 and 1.33, with varied y directions and alternating receiving ends. A match has five rallies. A rally draws with no point after 1600 ticks or when a post-return physical state repeats. More points wins the match.

Design a fixed neural network controller that wins as many matches as possible in the Pong variant below. Return the actual numeric weights and biases in one response.

Architecture: 6 → 12 → 4 → 3 → 3. Apply ReLU after every hidden layer; the three output scores are linear. Each dense layer computes output = W * input + b. Use only the listed layers. Return JSON {"layers":[{"weights":[[...]],"bias":[...]}, ...]}, ordered from input to output. Each weights matrix has one row per output neuron and one column per input to that layer. Include all coefficients and biases as finite numbers. Do not return code, formulas in place of numbers, or prose.

Rules:
- Coordinates are x along the length of a 360 by 160 court, and y across it. Each policy sees itself at the left end. The ball contacts the left paddle plane at x=15 and the right paddle plane at x=345. Paddle centers are limited to y=15..145. Each paddle has a half-width of 15 along y.
- The six inputs, in order, are (ballX-180)/180, (ballY-80)/80, ballVx/6, ballVy/6, (ownPaddleY-80)/80, (opponentPaddleY-80)/80. For the right player, reflect ballX about x=180 and negate ballVx before this normalization. Y and vy retain their signs. Negative observed vx means the ball is coming toward you.
- Both policies choose simultaneously from the same pre-tick state. Outputs in order mean move y by -2, 0, or +2. The largest score wins, with ties resolved in favor of the earliest output. Move both paddles and clamp their positions first, then advance the ball by (vx,vy).
- If the updated ball y is below 4, reflect it to 8-y. If it is above 156, reflect it to 312-y. In either case negate vy. Then check paddle contact when the ball has reached or crossed its destination plane.
- At contact let offset be updated ballY minus updated receiving paddleY. A miss (absolute offset greater than 15) ends the rally and awards one point to the other player. On a hit, snap x to the contact plane. Reverse vx with new magnitude s=min(13, abs(vx)+1). Set vy=round(9*s*offset/(15*13)), where half ties round toward positive infinity.
- Each rally starts with the ball at (180,80) and both paddles at y=80. The serve has abs(vx)=5 and abs(vy) between 0.89 and 1.33, with varied y directions and alternating receiving ends. A match has five rallies. A rally draws with no point after 1600 ticks or when a post-return physical state repeats. More points wins the match.

Design brief:
Allocate the neurons around a paddle target. When the ball approaches, estimate where it will reach your contact plane using its position and velocity, accounting for a wall bounce. When it departs, use a separate recovery target. On an incoming ball, use the opponent's position to choose a contact offset that makes the return harder to reach. Convert target error into the three action scores with a small stay zone.

Before emitting the weights, check the signs and scales on a straight incoming ball, a wall bounce, a departing ball, and the same incoming shot with the opponent on opposite sides. Return only the numeric matrices and biases.