Developed hand-designed 9x9 Go policy network

Summary

Closed, Completed

Build
Implemented initial hand-designed Go policy network.
Build
Verify
Ran multiple game passes to evaluate network performance.
Verify
Edit
Refined network design with safety and strategic improvements.
Edit
Verify
Validated final network performance against various opponents.
Verify

TASK. Write, from scratch, a policy network that plays 9 x 9 Go as strongly as possible. All of its weights start at zero; you set them. The goal is to win as many games as possible against the opponents described under HOW IT WILL BE MEASURED.

THE GAME: Go on a 9 x 9 board (the pgx implementation, "go_9x9"). Rows are numbered 0 (top) to 8 (bottom), columns 0 (left) to 8 (right). Black moves first; players alternate. A move either places a stone of the mover's colour on an empty point or passes. A group of stones with no adjacent empty point is captured and removed. Suicide is illegal, and so is retaking a single stone immediately (simple ko). A move that repeats an earlier whole-board position ends the game at once, and the player who made it loses. The game ends after two consecutive passes, or after 162 moves in total. Then it is scored by area: each player's stones on the board plus the empty points surrounded only by that player's stones; White receives a komi of 7.5 points. The higher score wins (no draws).

THE NETWORK. Input: 243 numbers, each 0 or 1: three planes over the 81 points, from the point of view of the player to move; input number (row * 9 + column) * 3 + plane, where plane 0 is 1 where the player to move has a stone, plane 1 is 1 where the opponent has a stone, and plane 2 is 1 everywhere when White is to move (0 when Black is). Output: 82 logits: output row * 9 + column = play at that point, output 81 = pass. h_1 = relu(x @ L1.W + L1.b) (512 units) h_k = relu(h_(k-1) @ Lk.W + Lk.b) for k = 2..6 (512 units each) logits = h_6 @ out.W + out.b (82) Tensors: L1.W (243 x 512), L1.b (512), L2.W .. L6.W (512 x 512), L2.b .. L6.b (512), out.W (512 x 82), out.b (82); 1,480,274 numbers. Row i of a weight matrix belongs to input (or unit) i, column j to unit (or output) j. Only the values can be set; the architecture is fixed. Illegal moves (occupied points, suicide, simple ko) are removed before every move, and the move is sampled from the softmax of the remaining logits (temperature 1). There is no search: the network's output is the whole policy.

HOW IT WILL BE MEASURED. After you finish, your network plays 2,000 games against each of several hidden opponents: policy networks trained with reinforcement learning on this game (self-play; they beat a uniform random player 79-98 % of the time), and a uniform random player. Your network is Black in half of the games; the moves of both players are sampled from their policies; there is no search. Score: wins / games; the main measure is the average over the trained opponents. You never see these opponents.

WHAT YOU HAVE.

A Python sandbox (no internet, no GPU, no Go library) with a working folder /work that is kept between runs. Available: numpy, scipy, scikit-learn, PyTorch (CPU), pandas, sympy, networkx. Before your code runs, these are defined: SHAPES and NAMES (the tensors), load(name) (a float32 numpy array: the values you saved, zeros at the start), save(name, array) (stores a tensor of the same shape), saved() (the names saved so far), inp(row, col, plane) (the input index), mv(row, col) (the output index) and PASS = 81. A run may take up to 900 seconds and 96 GB; its printed output (cut at 20,000 characters) is sent back to you. To run code, reply with exactly one block that starts with ```python and nothing else. You may run code up to 120 times.
Two pass tools with a budget of 25 passes in total; each use of either tool counts as one pass (a request that fails is not counted).

A games pass plays up to 128 games of your current network (the tensors saved so far) against practice opponents: reply with exactly one block that starts with ```games and contains JSON {"games": {"strong": 64, "weak": 32, "random": 32}} (even counts, at most 128 games in total; your network is Black in half of the games against each opponent). The practice opponents: "random" (uniform random legal moves), "weak" and "strong" (two policy networks trained with reinforcement learning on this game, not among the evaluation opponents; they beat a random player 95 % and 98 % of the time). You get a summary per opponent (score, wins, mean area margin, stones captured and lost, passes, game length, how the games ended) and, in /work/pass_/, every game's moves, your network's top-5 moves with their probabilities at each of its moves, and the board after every move.
A positions pass runs your current network on up to 256 positions: reply with exactly one block that starts with ```positions and contains JSON {"positions": [{"board": ["9 strings of 9 characters: X = Black, O = White, . = empty", ...], "to_move": "black" or "white"} or {"moves": [a list of moves from the empty board: row * 9 + column, or "pass"]}, ...], "activations": true or false}. You get the top 3 moves per position and, in /work/pass_/, the inputs, logits, probabilities over the legal moves and (if asked) the activations of every hidden layer. For "board" positions the legal moves are the empty points and pass (suicide and ko are not checked); for "moves" positions they are exact. The session continues until every pass is used. At the end, the tensors you saved with save() are your network; tensors never saved stay zero.

RULES. Not allowed: (a) gradient descent or any other training: no backward passes, autograd or optimizers, and no fitting of weights to data or labels (no regression or least squares on examples); (b) search: no game-tree search (MCTS, minimax, rollouts, looking ahead over sequences of moves), neither to choose moves nor to produce target moves or values for setting weights; (c) running the network outside the pass tools: do not compute your network's output or activations on any input in the sandbox; (d) internet access (there is none). Allowed: writing the weights by code from your own design (formulas, loops, symmetries, linear algebra on the weights themselves), Go-rule code in the sandbox (for example to work out the liberties of groups on boards you choose) as long as it does not run the network or search, and any analysis of the pass results.

HOW TO FINISH. When every pass is used and your network is saved, reply WITHOUT a code or pass block: describe its design and why you expect it to win.

python

import numpy as npimport os, json
os.makedirs("/work", exist_ok=True)
# All coefficients below are hand-designed, not learned.SOURCE = r'''import numpy as np
P = dict(    capture=7.0, rescue=4.5, self_atari=10.0,    eye=11.0, crowd=2.5, pass_bias=-2.0,    friend_adj=0.15, enemy_adj=0.35,    friend_diag=0.35, enemy_diag=0.10,    friend_jump=0.65, enemy_jump=0.15,    friend_knight=0.25, enemy_knight=0.05,    seed_shared_liberty=1.0,)P.update(overrides)
W = [np.zeros((243 if k == 0 else 512, 512), np.float32)     for k in range(6)]B = [np.zeros(512, np.float32) for _ in range(6)]OW = np.zeros((512, 82), np.float32)OB = np.zeros(82, np.float32)
def unit(ch, p):    return ch * 81 + p
def points(p, offsets):    r, c = divmod(p, 9)    return [(r+dr)*9+c+dc for dr,dc in offsets            if 0 <= r+dr < 9 and 0 <= c+dc < 9]
AD = [(1,0),(-1,0),(0,1),(0,-1)]DI = [(1,1),(1,-1),(-1,1),(-1,-1)]JU = [(2,0),(-2,0),(0,2),(0,-2)]KN = [(a,b) for a,b in      [(1,2),(1,-2),(-1,2),(-1,-2),(2,1),(2,-1),(-2,1),(-2,-1)]]adj = [points(p, AD) for p in range(81)]diag = [points(p, DI) for p in range(81)]
def add(k, dest, src, value):    W[k][src, dest] += value
def carry(k, ch, p):    add(k, unit(ch,p), unit(ch,p), 1)
# First layer:# own, enemy, and two ReLUs per colour encoding 0/1 local liberties.# D-Z is exactly 1 for a stone with at most one local liberty.for p in range(81):    add(0, unit(0,p), 3*p, 1)    add(0, unit(1,p), 3*p+1, 1)    for color, dch, zch in [(0,2,3),(1,4,5)]:        for ch, coefficient in [(dch,2),(zch,1)]:            dest = unit(ch,p)            add(0, dest, 3*p+color, coefficient)            B[0][dest] = -len(adj[p])            for q in adj[p]:                add(0, dest, 3*q, 1)                add(0, dest, 3*q+1, 1)
# Second layer adds conservative group-safety seeds:# a neighbouring stone with >=2 liberties makes the group safe;# adjacent stones each having a liberty provide distinct liberties.for p in range(81):    carry(1,0,p)    carry(1,1,p)    for color, dch, zch in [(0,2,3),(1,4,5)]:        for outch, subtract in [(dch,0),(zch,1)]:            dest = unit(outch,p)            add(1,dest,unit(dch,p),1)            B[1][dest] = -subtract            for q in adj[p]:                # -4 * (stone - D + Z)                add(1,dest,unit(color,q),-4)                add(1,dest,unit(dch,q),4)                add(1,dest,unit(zch,q),-4)                # D-2Z is 1 for exactly one local liberty.                add(1,dest,unit(dch,q),-P["seed_shared_liberty"])                add(1,dest,unit(zch,q),2*P["seed_shared_liberty"])
# Third layer caps danger at one and propagates group safety.# Channels become own, enemy, own-danger, enemy-danger, eye, bad-diagonal.for p in range(81):    carry(2,0,p)    carry(2,1,p)    for color,dch,zch,outch in [(0,2,3,2),(1,4,5,3)]:        dest = unit(outch,p)        add(2,dest,unit(dch,p),1)        add(2,dest,unit(zch,p),-1)        for q in adj[p]:            add(2,dest,unit(color,q),-1)            add(2,dest,unit(dch,q),1)            add(2,dest,unit(zch,q),-1)    dest = unit(4,p)    B[2][dest] = 1-len(adj[p])    for q in adj[p]:        add(2,dest,unit(0,q),1)    dest = unit(5,p)    # An interior eye tolerates one hostile diagonal; an edge eye none.    B[2][dest] = -1 if len(adj[p]) == 4 else 0    for q in diag[p]:        add(2,dest,unit(1,q),1)
# Fourth layer: more group-safety propagation and true-eye recognition.for p in range(81):    for ch in (0,1):        carry(3,ch,p)    for color,dch in [(0,2),(1,3)]:        dest = unit(dch,p)        add(3,dest,unit(dch,p),1)        for q in adj[p]:            add(3,dest,unit(color,q),-1)            add(3,dest,unit(dch,q),1)    add(3,unit(4,p),unit(4,p),1)    add(3,unit(4,p),unit(5,p),-1)    add(3,unit(5,p),unit(4,p),1)
# Fifth layer: final safety propagation, preserving eye features.for p in range(81):    for ch in (0,1,4,5):        carry(4,ch,p)    for color,dch in [(0,2),(1,3)]:        dest = unit(dch,p)        add(4,dest,unit(dch,p),1)        for q in adj[p]:            add(4,dest,unit(color,q),-1)            add(4,dest,unit(dch,q),1)
# Sixth layer computes action features.for p in range(81):    # Capture and rescue urgency.    for q in adj[p]:        add(5,unit(0,p),unit(3,q),1)        add(5,unit(1,p),unit(2,q),1)
    # Insufficient escape liberties, waived for captures.    dest = unit(2,p)    B[5][dest] = 2-len(adj[p])    for q in adj[p]:        add(5,dest,unit(0,q),1)        add(5,dest,unit(1,q),1)        add(5,dest,unit(0,q),-1.5)        add(5,dest,unit(2,q),1.5)        add(5,dest,unit(3,q),-2)
    add(5,unit(3,p),unit(4,p),1)
    # Discourage dense, redundant own-stone shapes.    dest = unit(4,p)    B[5][dest] = -2    for q in adj[p]:        add(5,dest,unit(0,q),1)    for q in diag[p]:        add(5,dest,unit(0,q),0.5)
    # A positive offset permits a signed linear positional feature.    dest = unit(5,p)    B[5][dest] = 20    for offsets, ownkey, enemykey in [        (AD,"friend_adj","enemy_adj"),        (DI,"friend_diag","enemy_diag"),        (JU,"friend_jump","enemy_jump"),        (KN,"friend_knight","enemy_knight"),    ]:        for q in points(p,offsets):            add(5,dest,unit(0,q),P[ownkey])            add(5,dest,unit(1,q),P[enemykey])    add(5,dest,unit(5,p),-1.5)
    for ch,coef in enumerate([        P["capture"], P["rescue"], -P["self_atari"],        -P["eye"], -P["crowd"], 1.0    ]):        OW[unit(ch,p),p] = coef
    r,c = divmod(p,9)    edge = min(r,c,8-r,8-c)    positional = [-0.9,0.1,0.75,0.65,0.65][edge]    if (r,c) in [(2,2),(2,6),(6,2),(6,6),(4,4)]:        positional += 0.35    OB[p] = -20 + positional
OB[81] = P["pass_bias"]
for k in range(6):    save("L%d.W" % (k+1), W[k])    save("L%d.b" % (k+1), B[k])save("out.W", OW)save("out.b", OB)'''open("/work/design.py", "w").write(SOURCE)overrides = {}exec(SOURCE)print("Saved all 14 tensors. Hand-built liberty, capture, rescue, eye, and shape policy.")print("Parameters:", P)print("Saved tensors:", saved())

import numpy as npimport os, json
os.makedirs("/work", exist_ok=True)
# All coefficients below are hand-designed, not learned.SOURCE = r'''import numpy as np
P = dict(    capture=7.0, rescue=4.5, self_atari=10.0,    eye=11.0, crowd=2.5, pass_bias=-2.0,    friend_adj=0.15, enemy_adj=0.35,    friend_diag=0.35, enemy_diag=0.10,    friend_jump=0.65, enemy_jump=0.15,    friend_knight=0.25, enemy_knight=0.05,    seed_shared_liberty=1.0,)P.update(overrides)
W = [np.zeros((243 if k == 0 else 512, 512), np.float32)     for k in range(6)]B = [np.zeros(512, np.float32) for _ in range(6)]OW = np.zeros((512, 82), np.float32)OB = np.zeros(82, np.float32)
def unit(ch, p):    return ch * 81 + p
def points(p, offsets):    r, c = divmod(p, 9)    return [(r+dr)*9+c+dc for dr,dc in offsets            if 0 <= r+dr < 9 and 0 <= c+dc < 9]
AD = [(1,0),(-1,0),(0,1),(0,-1)]DI = [(1,1),(1,-1),(-1,1),(-1,-1)]JU = [(2,0),(-2,0),(0,2),(0,-2)]KN = [(a,b) for a,b in      [(1,2),(1,-2),(-1,2),(-1,-2),(2,1),(2,-1),(-2,1),(-2,-1)]]adj = [points(p, AD) for p in range(81)]diag = [points(p, DI) for p in range(81)]
def add(k, dest, src, value):    W[k][src, dest] += value
def carry(k, ch, p):    add(k, unit(ch,p), unit(ch,p), 1)
# First layer:# own, enemy, and two ReLUs per colour encoding 0/1 local liberties.# D-Z is exactly 1 for a stone with at most one local liberty.for p in range(81):    add(0, unit(0,p), 3*p, 1)    add(0, unit(1,p), 3*p+1, 1)    for color, dch, zch in [(0,2,3),(1,4,5)]:        for ch, coefficient in [(dch,2),(zch,1)]:            dest = unit(ch,p)            add(0, dest, 3*p+color, coefficient)            B[0][dest] = -len(adj[p])            for q in adj[p]:                add(0, dest, 3*q, 1)                add(0, dest, 3*q+1, 1)
# Second layer adds conservative group-safety seeds:# a neighbouring stone with >=2 liberties makes the group safe;# adjacent stones each having a liberty provide distinct liberties.for p in range(81):    carry(1,0,p)    carry(1,1,p)    for color, dch, zch in [(0,2,3),(1,4,5)]:        for outch, subtract in [(dch,0),(zch,1)]:            dest = unit(outch,p)            add(1,dest,unit(dch,p),1)            B[1][dest] = -subtract            for q in adj[p]:                # -4 * (stone - D + Z)                add(1,dest,unit(color,q),-4)                add(1,dest,unit(dch,q),4)                add(1,dest,unit(zch,q),-4)                # D-2Z is 1 for exactly one local liberty.                add(1,dest,unit(dch,q),-P["seed_shared_liberty"])                add(1,dest,unit(zch,q),2*P["seed_shared_liberty"])
# Third layer caps danger at one and propagates group safety.# Channels become own, enemy, own-danger, enemy-danger, eye, bad-diagonal.for p in range(81):    carry(2,0,p)    carry(2,1,p)    for color,dch,zch,outch in [(0,2,3,2),(1,4,5,3)]:        dest = unit(outch,p)        add(2,dest,unit(dch,p),1)        add(2,dest,unit(zch,p),-1)        for q in adj[p]:            add(2,dest,unit(color,q),-1)            add(2,dest,unit(dch,q),1)            add(2,dest,unit(zch,q),-1)    dest = unit(4,p)    B[2][dest] = 1-len(adj[p])    for q in adj[p]:        add(2,dest,unit(0,q),1)    dest = unit(5,p)    # An interior eye tolerates one hostile diagonal; an edge eye none.    B[2][dest] = -1 if len(adj[p]) == 4 else 0    for q in diag[p]:        add(2,dest,unit(1,q),1)
# Fourth layer: more group-safety propagation and true-eye recognition.for p in range(81):    for ch in (0,1):        carry(3,ch,p)    for color,dch in [(0,2),(1,3)]:        dest = unit(dch,p)        add(3,dest,unit(dch,p),1)        for q in adj[p]:            add(3,dest,unit(color,q),-1)            add(3,dest,unit(dch,q),1)    add(3,unit(4,p),unit(4,p),1)    add(3,unit(4,p),unit(5,p),-1)    add(3,unit(5,p),unit(4,p),1)
# Fifth layer: final safety propagation, preserving eye features.for p in range(81):    for ch in (0,1,4,5):        carry(4,ch,p)    for color,dch in [(0,2),(1,3)]:        dest = unit(dch,p)        add(4,dest,unit(dch,p),1)        for q in adj[p]:            add(4,dest,unit(color,q),-1)            add(4,dest,unit(dch,q),1)
# Sixth layer computes action features.for p in range(81):    # Capture and rescue urgency.    for q in adj[p]:        add(5,unit(0,p),unit(3,q),1)        add(5,unit(1,p),unit(2,q),1)
    # Insufficient escape liberties, waived for captures.    dest = unit(2,p)    B[5][dest] = 2-len(adj[p])    for q in adj[p]:        add(5,dest,unit(0,q),1)        add(5,dest,unit(1,q),1)        add(5,dest,unit(0,q),-1.5)        add(5,dest,unit(2,q),1.5)        add(5,dest,unit(3,q),-2)
    add(5,unit(3,p),unit(4,p),1)
    # Discourage dense, redundant own-stone shapes.    dest = unit(4,p)    B[5][dest] = -2    for q in adj[p]:        add(5,dest,unit(0,q),1)    for q in diag[p]:        add(5,dest,unit(0,q),0.5)
    # A positive offset permits a signed linear positional feature.    dest = unit(5,p)    B[5][dest] = 20    for offsets, ownkey, enemykey in [        (AD,"friend_adj","enemy_adj"),        (DI,"friend_diag","enemy_diag"),        (JU,"friend_jump","enemy_jump"),        (KN,"friend_knight","enemy_knight"),    ]:        for q in points(p,offsets):            add(5,dest,unit(0,q),P[ownkey])            add(5,dest,unit(1,q),P[enemykey])    add(5,dest,unit(5,p),-1.5)
    for ch,coef in enumerate([        P["capture"], P["rescue"], -P["self_atari"],        -P["eye"], -P["crowd"], 1.0    ]):        OW[unit(ch,p),p] = coef
    r,c = divmod(p,9)    edge = min(r,c,8-r,8-c)    positional = [-0.9,0.1,0.75,0.65,0.65][edge]    if (r,c) in [(2,2),(2,6),(6,2),(6,6),(4,4)]:        positional += 0.35    OB[p] = -20 + positional
OB[81] = P["pass_bias"]
for k in range(6):    save("L%d.W" % (k+1), W[k])    save("L%d.b" % (k+1), B[k])save("out.W", OW)save("out.b", OB)'''open("/work/design.py", "w").write(SOURCE)overrides = {}exec(SOURCE)print("Saved all 14 tensors. Hand-built liberty, capture, rescue, eye, and shape policy.")print("Parameters:", P)print("Saved tensors:", saved())

OUTPUT of Python run 1 of at most 120:

Saved all 14 tensors. Hand-built liberty, capture, rescue, eye, and shape policy.Parameters: {'capture': 7.0, 'rescue': 4.5, 'self_atari': 10.0, 'eye': 11.0, 'crowd': 2.5, 'pass_bias': -2.0, 'friend_adj': 0.15, 'enemy_adj': 0.35, 'friend_diag': 0.35, 'enemy_diag': 0.1, 'friend_jump': 0.65, 'enemy_jump': 0.15, 'friend_knight': 0.25, 'enemy_knight': 0.05, 'seed_shared_liberty': 1.0}Saved tensors: ['L1.W', 'L1.b', 'L2.W', 'L2.b', 'L3.W', 'L3.b', 'L4.W', 'L4.b', 'L5.W', 'L5.b', 'L6.W', 'L6.b', 'out.W', 'out.b']

Saved all 14 tensors. Hand-built liberty, capture, rescue, eye, and shape policy.Parameters: {'capture': 7.0, 'rescue': 4.5, 'self_atari': 10.0, 'eye': 11.0, 'crowd': 2.5, 'pass_bias': -2.0, 'friend_adj': 0.15, 'enemy_adj': 0.35, 'friend_diag': 0.35, 'enemy_diag': 0.1, 'friend_jump': 0.65, 'enemy_jump': 0.15, 'friend_knight': 0.25, 'enemy_knight': 0.05, 'seed_shared_liberty': 1.0}Saved tensors: ['L1.W', 'L1.b', 'L2.W', 'L2.b', 'L3.W', 'L3.b', 'L4.W', 'L4.b', 'L5.W', 'L5.b', 'L6.W', 'L6.b', 'out.W', 'out.b']

Games pass 1 of 25 done.

{"ok": true, "pass": 1, "kind": "games", "per_opponent": {"random": {"games": 32, "score": 1.0, "wins": 32, "losses": 0, "score_as_black": 1.0, "score_as_white": 1.0, "mean_margin": 57.65625, "mean_length": 104.6875, "stones_captured_per_game": 41.75, "stones_lost_per_game": 0.84375, "passes_per_game": 1.1875, "games_with_a_pass_before_move_60": 0, "end_two_passes": 32, "end_repetition": 0, "end_move_limit": 0}, "strong": {"games": 64, "score": 0.375, "wins": 24, "losses": 40, "score_as_black": 0.3125, "score_as_white": 0.4375, "mean_margin": 6.25, "mean_length": 103.8125, "stones_captured_per_game": 18.640625, "stones_lost_per_game": 20.140625, "passes_per_game": 2.796875, "games_with_a_pass_before_move_60": 2, "end_two_passes": 59, "end_repetition": 5, "end_move_limit": 0}, "weak": {"games": 32, "score": 0.65625, "wins": 21, "losses": 11, "score_as_black": 0.625, "score_as_white": 0.6875, "mean_margin": 31.96875, "mean_length": 109.03125, "stones_captured_per_game": 31.4375, "stones_lost_per_game": 14.34375, "passes_per_game": 2.40625, "games_with_a_pass_before_move_60": 0, "end_two_passes": 31, "end_repetition": 0, "end_move_limit": 1}}, "files": "/work/pass_1/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 1, "kind": "games", "per_opponent": {"random": {"games": 32, "score": 1.0, "wins": 32, "losses": 0, "score_as_black": 1.0, "score_as_white": 1.0, "mean_margin": 57.65625, "mean_length": 104.6875, "stones_captured_per_game": 41.75, "stones_lost_per_game": 0.84375, "passes_per_game": 1.1875, "games_with_a_pass_before_move_60": 0, "end_two_passes": 32, "end_repetition": 0, "end_move_limit": 0}, "strong": {"games": 64, "score": 0.375, "wins": 24, "losses": 40, "score_as_black": 0.3125, "score_as_white": 0.4375, "mean_margin": 6.25, "mean_length": 103.8125, "stones_captured_per_game": 18.640625, "stones_lost_per_game": 20.140625, "passes_per_game": 2.796875, "games_with_a_pass_before_move_60": 2, "end_two_passes": 59, "end_repetition": 5, "end_move_limit": 0}, "weak": {"games": 32, "score": 0.65625, "wins": 21, "losses": 11, "score_as_black": 0.625, "score_as_white": 0.6875, "mean_margin": 31.96875, "mean_length": 109.03125, "stones_captured_per_game": 31.4375, "stones_lost_per_game": 14.34375, "passes_per_game": 2.40625, "games_with_a_pass_before_move_60": 0, "end_two_passes": 31, "end_repetition": 0, "end_move_limit": 1}}, "files": "/work/pass_1/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 24.

python

import json, numpy as npg=json.load(open("/work/pass_1/games.json"))print(type(g))if isinstance(g,dict):    print(g.keys())else:    print("Games",len(g))    print(json.dumps(g[0],indent=1)[:10000])

import json, numpy as npg=json.load(open("/work/pass_1/games.json"))print(type(g))if isinstance(g,dict):    print(g.keys())else:    print("Games",len(g))    print(json.dumps(g[0],indent=1)[:10000])

OUTPUT of Python run 2 of at most 120:

<class 'list'>Games 128{ "game": 0, "opponent": "random", "net_color": "black", "result": 1, "margin": 73.5, "end": "two passes", "stones_captured_by_net": 69, "stones_lost_by_net": 0, "moves": [  {   "ply": 0,   "by": "net",   "move": 67,   "top5": [    [     20,     0.0316    ],    [     24,     0.0316    ],    [     56,     0.0316    ],    [     60,     0.0316    ],    [     40,     0.0286    ]   ]  },  {   "ply": 1,   "by": "opponent",   "move": 65  },  {   "ply": 2,   "by": "net",   "move": 57,   "top5": [    [     56,     0.0518    ],    [     60,     0.0365    ],    [     49,     0.0347    ],    [     57,     0.0314    ],    [     20,     0.0284    ]   ]  },  {   "ply": 3,   "by": "opponent",   "move": 3  },  {   "ply": 4,   "by": "net",   "move": 24,   "top5": [    [     56,     0.0538    ],    [     59,     0.0487    ],    [     49,     0.044    ],    [     60,     0.0326    ],    [     39,     0.031    ]   ]  },  {   "ply": 5,   "by": "opponent",   "move": 59  },  {   "ply": 6,   "by": "net",   "move": 29,   "top5": [    [     56,     0.05    ],    [     49,     0.0453    ],    [     60,     0.0431    ],    [     50,     0.0353    ],    [     22,     0.0335    ]   ]  },  {   "ply": 7,   "by": "opponent",   "move": 63  },  {   "ply": 8,   "by": "net",   "move": 20,   "top5": [    [     47,     0.047    ],    [     56,     0.047    ],    [     49,     0.0404    ],    [     22,     0.0384    ],    [     60,     0.0384    ]   ]  },  {   "ply": 9,   "by": "opponent",   "move": 4  },  {   "ply": 10,   "by": "net",   "move": 41,   "top5": [    [     22,     0.0764    ],    [     39,     0.0419    ],    [     47,     0.0419    ],    [     56,     0.0419    ],    [     31,     0.0379    ]   ]  },  {   "ply": 11,   "by": "opponent",   "move": 72  },  {   "ply": 12,   "by": "net",   "move": 39,   "top5": [    [     22,     0.0841    ],    [     39,     0.0689    ],    [     31,     0.0462    ],    [     49,     0.0439    ],    [     60,     0.0378    ]   ]  },  {   "ply": 13,   "by": "opponent",   "move": 0  },  {   "ply": 14,   "by": "net",   "move": 50,   "top5": [    [     22,     0.1    ],    [     31,     0.0606    ],    [     49,     0.0577    ],    [     47,     0.0472    ],    [     56,     0.0427    ]   ]  },  {   "ply": 15,   "by": "opponent",   "move": 17  },  {   "ply": 16,   "by": "net",   "move": 22,   "top5": [    [     22,     0.0985    ],    [     31,     0.0767    ],    [     48,     0.0568    ],    [     60,     0.0489    ],    [     47,     0.0465    ]   ]  },  {   "ply": 17,   "by": "opponent",   "move": 42  },  {   "ply": 18,   "by": "net",   "move": 31,   "top5": [    [     31,     0.0974    ],    [     32,     0.0621    ],    [     60,     0.0591    ],    [     48,     0.0591    ],    [     47,     0.0484    ]   ]  },  {   "ply": 19,   "by": "opponent",   "move": 34  },  {   "ply": 20,   "by": "net",   "move": 48,   "top5": [    [     33,     0.1196    ],    [     48,     0.0802    ],    [     60,     0.0625    ],    [     47,     0.0511    ],    [     38,     0.0486    ]   ]  },  {   "ply": 21,   "by": "opponent",   "move": 44  },  {   "ply": 22,   "by": "net",   "move": 47,   "top5": [    [     33,     0.1443    ],    [     56,     0.0753    ],    [     60,     0.0717    ],    [     47,     0.0682    ],    [     13,     0.0393    ]   ]  },  {   "ply": 23,   "by": "opponent",   "move": 10  },  {   "ply": 24,   "by": "net",   "move": 79,   "top5": [    [     33,     0.1571    ],    [     60,     0.078    ],    [     37,     0.0473    ],    [     12,     0.045    ],    [     13,     0.0428    ]   ]  },  {   "ply": 25,   "by": "opponent",   "move": 16  },  {   "ply": 26,   "by": "net",   "move": 60,   "top5": [    [     33,     0.1548    ],    [     60,     0.0939    ],    [     37,     0.0443    ],    [     12,     0.0422    ],    [     13,     0.0401    ]   ]  },  {   "ply": 27,   "by": "opponent",   "move": 18  },  {   "ply": 28,   "by": "net",   "move": 15,   "top5": [    [     33,     0.1629    ],    [     68,     0.0542    ],    [     37,     0.0491    ],    [     12,     0.0444    ],    [     13,     0.0422    ]   ]  },  {   "ply": 29,   "by": "opponent",   "move": 13  },  {   "ply": 30,   "by": "net",   "move": 69,   "top5": [    [     33,     0.2731    ],    [     12,     0.0551    ],    [     68,     0.0475    ],    [     37,     0.0429    ],    [     11,     0.0389    ]   ]  },  {   "ply": 31,   "by": "opponent",   "move": 66  },  {   "ply": 32,   "by": "net",   "move": 56,   "top5": [    [     33,     0.2826    ],    [     12,     0.0571    ],    [     37,     0.0444    ],    [     52,     0.0444    ],    [     11,     0.0402    ]   ]  },  {   "ply": 33,   "by": "opponent",   "move": 28  },  {   "ply": 34,   "by": "net",   "move": 55,   "top5": [    [     33,     0.2709    ],    [     37,     0.0776    ],    [     12,     0.0547    ],    [     46,     0.0547    ],    [     58,     0.0448    ]   ]  },  {   "ply": 35,   "by": "opponent",   "move": 71  },  {   "ply": 36,   "by": "net",   "move": 12,   "top5": [    [     33,     0.265    ],    [     37,     0.1454    ],    [     12,     0.0535    ],    [     52,     0.0438    ],    [     64,     0.0438    ]   ]  },  {   "ply": 37,   "by": "opponent",   "move": 36  },  {   "ply": 38,   "by": "net",   "move": 37,   "top5": [    [     33,     0.2518    ],    [     37,     0.1961    ],    [     14,     0.0653    ],    [     11,     0.0438    ],    [     52,     0.0416    ]   ]  },  {   "ply": 39,   "by": "opponent",   "move": 40  },  {   "ply": 40,   "by": "net",   "move": 49,   "top5": [    [     49,     0.9625    ],    [     33,     0.0124    ],    [     14,     0.0031    ],    [     58,     0.0023    ],    [     11,     0.0021    ]   ]  },  {   "ply": 41,   "by": "opponent",   "move": 73  },  {   "ply": 42,   "by": "net",   "move": 33,   "top5": [    [     33,     0.3596    ],    [     14,     0.0932    ],    [     11,     0.0625    ],    [     52,     0.0594    ],    [     51,     0.0565    ]   ]  },  {   "ply": 43,   "by": "opponent",   "move": 9  },  {   "ply": 44,   "by": "net",   "move": 6,   "top5": [    [     14,     0.1613    ],    [     51,     0.1459    ],    [     52,     0.1028    ],    [     11,     0.0978    ],    [     25,     0.0656    ]   ]  },  {   "ply": 45,   "by": "opponent",   "move": 23  },  {   "ply": 46,   "by": "net",   "move": 11,   "top5": [    [     51,     0.1527    ],    [     52,     0.1076    ],    [     11,     0.1023    ],    [     25,     0.1023    ],    [     14,     0.0973    ]   ]  },  {   "ply": 47,   "by": "opponent",   "move": 62  },  {   "ply": 48,   "by": "net",   "move": 14,   "top5": [    [     51,     0.1746    ],    [     52,     0.1294    ],    [     25,     0.1113    ],    [     14,     0.1059    ],    [     61,     0.1059    ]   ]  },  {   "ply": 49,   "by": "opponent",   "move": 80  },  {   "ply": 50,   "by": "net",   "move": 32,   "top5": [    [     32,     0.8954    ],    [     51,     0.02    ],    [     25,     0.0164    ],    [     52,     0.0148    ],    [     61,     0.0128    ]   ]  },  {   "ply": 51,   "by": "opponent",   "move": 52  },  {   "ply": 52,   "by": "net",   "move": 25,   "top5": [    [     51,     0.32    ],    [     25,     0.1589    ],    [     61,     0.1589    ],    [     68,     0.0479    ],    [     45,     0.0412    ]   ]  },  {   "ply": 53,   "by": "opponent",   "move": 45  },  {   "ply": 54,   "by": "net",   "move": 77,   "top5": [    [     51,     0.4059    ],    [     61,     0.2016    ],    [     68,     0.0607    ],    [     7,     0.0497    ],    [     2,     0.045    ]   ]  },  {   "ply": 55,   "by": "opponent",   "move": 26  },  {   "ply": 56,   "by": "net",   "move": 2,   "top5": [    [     51,     0.458    ],    [     61,     0.2275    ],    [     75,     0.0837    ],    [     7,     0.059    ],    [     2,     0.0508    ]   ]  },  {   "ply": 57,   "by": "opponent",   "move": 7  },  {   "ply": 58,   "by": "net",   "move": 5,   "top5": [    [     5,     0.9569    ],    [     51,     0.0225    ],    [     61,     0.0112    ],    [     75,     0.0041    ],    [     70,     0.0014    ]   ]  },  {   "ply": 59,   "by": "opponent",   "move": 1  },  {   "ply": 60,   "by": "net",   "move": 4,   "top5": [    [     51,     0.4149    ],    [     61,     0.206    ],    [     4,     0.1687    ],    [     75,     0.0758    ],    [     70,     0.0252    ]   ]  },  {   "ply": 61,   "by": "opponent",   "move": 8  },  {   "ply": 62,   "by": "net",   "move": 35,   "top5": [    [     35,     0.9914    ],    [     51,     0.0045    ],    [     61,     0.0022    ],    [     75,     0.0008    ],    [     70,     0.0003    ]   ]  },  {   "ply": 63,   "by": "opponent",   "move": 19  },  {   "ply": 64,   "by": "net",   "move": 43,   "top5": [    [     43,     0.9975    ],    [     26,     0.0019    ],    [     51,     0.0003    ],    [     61,     0.0001    ],    [     75,     0.0    ]   ]  },  {   "ply": 65,   "by": "opponent",   "move": 53  },

<class 'list'>Games 128{ "game": 0, "opponent": "random", "net_color": "black", "result": 1, "margin": 73.5, "end": "two passes", "stones_captured_by_net": 69, "stones_lost_by_net": 0, "moves": [  {   "ply": 0,   "by": "net",   "move": 67,   "top5": [    [     20,     0.0316    ],    [     24,     0.0316    ],    [     56,     0.0316    ],    [     60,     0.0316    ],    [     40,     0.0286    ]   ]  },  {   "ply": 1,   "by": "opponent",   "move": 65  },  {   "ply": 2,   "by": "net",   "move": 57,   "top5": [    [     56,     0.0518    ],    [     60,     0.0365    ],    [     49,     0.0347    ],    [     57,     0.0314    ],    [     20,     0.0284    ]   ]  },  {   "ply": 3,   "by": "opponent",   "move": 3  },  {   "ply": 4,   "by": "net",   "move": 24,   "top5": [    [     56,     0.0538    ],    [     59,     0.0487    ],    [     49,     0.044    ],    [     60,     0.0326    ],    [     39,     0.031    ]   ]  },  {   "ply": 5,   "by": "opponent",   "move": 59  },  {   "ply": 6,   "by": "net",   "move": 29,   "top5": [    [     56,     0.05    ],    [     49,     0.0453    ],    [     60,     0.0431    ],    [     50,     0.0353    ],    [     22,     0.0335    ]   ]  },  {   "ply": 7,   "by": "opponent",   "move": 63  },  {   "ply": 8,   "by": "net",   "move": 20,   "top5": [    [     47,     0.047    ],    [     56,     0.047    ],    [     49,     0.0404    ],    [     22,     0.0384    ],    [     60,     0.0384    ]   ]  },  {   "ply": 9,   "by": "opponent",   "move": 4  },  {   "ply": 10,   "by": "net",   "move": 41,   "top5": [    [     22,     0.0764    ],    [     39,     0.0419    ],    [     47,     0.0419    ],    [     56,     0.0419    ],    [     31,     0.0379    ]   ]  },  {   "ply": 11,   "by": "opponent",   "move": 72  },  {   "ply": 12,   "by": "net",   "move": 39,   "top5": [    [     22,     0.0841    ],    [     39,     0.0689    ],    [     31,     0.0462    ],    [     49,     0.0439    ],    [     60,     0.0378    ]   ]  },  {   "ply": 13,   "by": "opponent",   "move": 0  },  {   "ply": 14,   "by": "net",   "move": 50,   "top5": [    [     22,     0.1    ],    [     31,     0.0606    ],    [     49,     0.0577    ],    [     47,     0.0472    ],    [     56,     0.0427    ]   ]  },  {   "ply": 15,   "by": "opponent",   "move": 17  },  {   "ply": 16,   "by": "net",   "move": 22,   "top5": [    [     22,     0.0985    ],    [     31,     0.0767    ],    [     48,     0.0568    ],    [     60,     0.0489    ],    [     47,     0.0465    ]   ]  },  {   "ply": 17,   "by": "opponent",   "move": 42  },  {   "ply": 18,   "by": "net",   "move": 31,   "top5": [    [     31,     0.0974    ],    [     32,     0.0621    ],    [     60,     0.0591    ],    [     48,     0.0591    ],    [     47,     0.0484    ]   ]  },  {   "ply": 19,   "by": "opponent",   "move": 34  },  {   "ply": 20,   "by": "net",   "move": 48,   "top5": [    [     33,     0.1196    ],    [     48,     0.0802    ],    [     60,     0.0625    ],    [     47,     0.0511    ],    [     38,     0.0486    ]   ]  },  {   "ply": 21,   "by": "opponent",   "move": 44  },  {   "ply": 22,   "by": "net",   "move": 47,   "top5": [    [     33,     0.1443    ],    [     56,     0.0753    ],    [     60,     0.0717    ],    [     47,     0.0682    ],    [     13,     0.0393    ]   ]  },  {   "ply": 23,   "by": "opponent",   "move": 10  },  {   "ply": 24,   "by": "net",   "move": 79,   "top5": [    [     33,     0.1571    ],    [     60,     0.078    ],    [     37,     0.0473    ],    [     12,     0.045    ],    [     13,     0.0428    ]   ]  },  {   "ply": 25,   "by": "opponent",   "move": 16  },  {   "ply": 26,   "by": "net",   "move": 60,   "top5": [    [     33,     0.1548    ],    [     60,     0.0939    ],    [     37,     0.0443    ],    [     12,     0.0422    ],    [     13,     0.0401    ]   ]  },  {   "ply": 27,   "by": "opponent",   "move": 18  },  {   "ply": 28,   "by": "net",   "move": 15,   "top5": [    [     33,     0.1629    ],    [     68,     0.0542    ],    [     37,     0.0491    ],    [     12,     0.0444    ],    [     13,     0.0422    ]   ]  },  {   "ply": 29,   "by": "opponent",   "move": 13  },  {   "ply": 30,   "by": "net",   "move": 69,   "top5": [    [     33,     0.2731    ],    [     12,     0.0551    ],    [     68,     0.0475    ],    [     37,     0.0429    ],    [     11,     0.0389    ]   ]  },  {   "ply": 31,   "by": "opponent",   "move": 66  },  {   "ply": 32,   "by": "net",   "move": 56,   "top5": [    [     33,     0.2826    ],    [     12,     0.0571    ],    [     37,     0.0444    ],    [     52,     0.0444    ],    [     11,     0.0402    ]   ]  },  {   "ply": 33,   "by": "opponent",   "move": 28  },  {   "ply": 34,   "by": "net",   "move": 55,   "top5": [    [     33,     0.2709    ],    [     37,     0.0776    ],    [     12,     0.0547    ],    [     46,     0.0547    ],    [     58,     0.0448    ]   ]  },  {   "ply": 35,   "by": "opponent",   "move": 71  },  {   "ply": 36,   "by": "net",   "move": 12,   "top5": [    [     33,     0.265    ],    [     37,     0.1454    ],    [     12,     0.0535    ],    [     52,     0.0438    ],    [     64,     0.0438    ]   ]  },  {   "ply": 37,   "by": "opponent",   "move": 36  },  {   "ply": 38,   "by": "net",   "move": 37,   "top5": [    [     33,     0.2518    ],    [     37,     0.1961    ],    [     14,     0.0653    ],    [     11,     0.0438    ],    [     52,     0.0416    ]   ]  },  {   "ply": 39,   "by": "opponent",   "move": 40  },  {   "ply": 40,   "by": "net",   "move": 49,   "top5": [    [     49,     0.9625    ],    [     33,     0.0124    ],    [     14,     0.0031    ],    [     58,     0.0023    ],    [     11,     0.0021    ]   ]  },  {   "ply": 41,   "by": "opponent",   "move": 73  },  {   "ply": 42,   "by": "net",   "move": 33,   "top5": [    [     33,     0.3596    ],    [     14,     0.0932    ],    [     11,     0.0625    ],    [     52,     0.0594    ],    [     51,     0.0565    ]   ]  },  {   "ply": 43,   "by": "opponent",   "move": 9  },  {   "ply": 44,   "by": "net",   "move": 6,   "top5": [    [     14,     0.1613    ],    [     51,     0.1459    ],    [     52,     0.1028    ],    [     11,     0.0978    ],    [     25,     0.0656    ]   ]  },  {   "ply": 45,   "by": "opponent",   "move": 23  },  {   "ply": 46,   "by": "net",   "move": 11,   "top5": [    [     51,     0.1527    ],    [     52,     0.1076    ],    [     11,     0.1023    ],    [     25,     0.1023    ],    [     14,     0.0973    ]   ]  },  {   "ply": 47,   "by": "opponent",   "move": 62  },  {   "ply": 48,   "by": "net",   "move": 14,   "top5": [    [     51,     0.1746    ],    [     52,     0.1294    ],    [     25,     0.1113    ],    [     14,     0.1059    ],    [     61,     0.1059    ]   ]  },  {   "ply": 49,   "by": "opponent",   "move": 80  },  {   "ply": 50,   "by": "net",   "move": 32,   "top5": [    [     32,     0.8954    ],    [     51,     0.02    ],    [     25,     0.0164    ],    [     52,     0.0148    ],    [     61,     0.0128    ]   ]  },  {   "ply": 51,   "by": "opponent",   "move": 52  },  {   "ply": 52,   "by": "net",   "move": 25,   "top5": [    [     51,     0.32    ],    [     25,     0.1589    ],    [     61,     0.1589    ],    [     68,     0.0479    ],    [     45,     0.0412    ]   ]  },  {   "ply": 53,   "by": "opponent",   "move": 45  },  {   "ply": 54,   "by": "net",   "move": 77,   "top5": [    [     51,     0.4059    ],    [     61,     0.2016    ],    [     68,     0.0607    ],    [     7,     0.0497    ],    [     2,     0.045    ]   ]  },  {   "ply": 55,   "by": "opponent",   "move": 26  },  {   "ply": 56,   "by": "net",   "move": 2,   "top5": [    [     51,     0.458    ],    [     61,     0.2275    ],    [     75,     0.0837    ],    [     7,     0.059    ],    [     2,     0.0508    ]   ]  },  {   "ply": 57,   "by": "opponent",   "move": 7  },  {   "ply": 58,   "by": "net",   "move": 5,   "top5": [    [     5,     0.9569    ],    [     51,     0.0225    ],    [     61,     0.0112    ],    [     75,     0.0041    ],    [     70,     0.0014    ]   ]  },  {   "ply": 59,   "by": "opponent",   "move": 1  },  {   "ply": 60,   "by": "net",   "move": 4,   "top5": [    [     51,     0.4149    ],    [     61,     0.206    ],    [     4,     0.1687    ],    [     75,     0.0758    ],    [     70,     0.0252    ]   ]  },  {   "ply": 61,   "by": "opponent",   "move": 8  },  {   "ply": 62,   "by": "net",   "move": 35,   "top5": [    [     35,     0.9914    ],    [     51,     0.0045    ],    [     61,     0.0022    ],    [     75,     0.0008    ],    [     70,     0.0003    ]   ]  },  {   "ply": 63,   "by": "opponent",   "move": 19  },  {   "ply": 64,   "by": "net",   "move": 43,   "top5": [    [     43,     0.9975    ],    [     26,     0.0019    ],    [     51,     0.0003    ],    [     61,     0.0001    ],    [     75,     0.0    ]   ]  },  {   "ply": 65,   "by": "opponent",   "move": 53  },

python

import json, numpy as npgames=json.load(open("/work/pass_1/games.json"))boards=np.load("/work/pass_1/boards.npy")AD=[[(r+dr)*9+c+dc for dr,dc in [(1,0),(-1,0),(0,1),(0,-1)]     if 0<=r+dr<9 and 0<=c+dc<9] for r in range(9) for c in range(9)]def groups(b):    seen=set(); out=[]    for p in range(81):        if not b[p] or p in seen: continue        col=int(b[p]); stack=[p]; stones=set([p]); libs=set(); seen.add(p)        while stack:            u=stack.pop()            for v in AD[u]:                if b[v]==0: libs.add(v)                elif b[v]==col and v not in seen:                    seen.add(v); stones.add(v); stack.append(v)        out.append((col,stones,libs))    return outdef show(b):    return "\n".join("".join(".XO"[{0:0,1:1,-1:2}[int(v)]] for v in b[r*9:(r+1)*9]) for r in range(9))open("/work/analysis_helpers.py","w").write('import numpy as np\nAD='+repr(AD)+'\n'+    '''def groups(b):    seen=set(); out=[]    for p in range(81):        if not b[p] or p in seen: continue        col=int(b[p]); stack=[p]; stones={p}; libs=set(); seen.add(p)        while stack:            u=stack.pop()            for v in AD[u]:                if b[v]==0: libs.add(v)                elif b[v]==col and v not in seen:                    seen.add(v); stones.add(v); stack.append(v)        out.append((col,stones,libs))    return outdef show(b):    return "\\n".join("".join({0:".",1:"X",-1:"O"}[int(v)] for v in b[r*9:(r+1)*9]) for r in range(9))''')losses=[(i,g) for i,g in enumerate(games) if g["opponent"]=="strong" and g["result"]<0]for i,g in losses[:5]:    print("\nGAME",i,"net",g["net_color"],"margin",g["margin"],"end",g["end"])    col=1 if g["net_color"]=="black" else -1    events=[]    for t,m in enumerate(g["moves"]):        if t==0 or m["by"]!="opponent": continue        lost=int(np.sum(boards[i,t-1]==col)-np.sum(boards[i,t]==col))        if lost>=4:            events.append((lost,t))    for lost,t in sorted(events,reverse=True)[:2]:        print("LOST",lost,"at",t,"opponent move",g["moves"][t]["move"])        s=max(0,t-3)        print("Board before own move",s+1,"\n"+show(boards[i,s]))        print("Sequence",g["moves"][s+1:t+1])        print("Own groups then",[(len(st),sorted(li)) for c,st,li in groups(boards[i,s]) if c==col])

import json, numpy as npgames=json.load(open("/work/pass_1/games.json"))boards=np.load("/work/pass_1/boards.npy")AD=[[(r+dr)*9+c+dc for dr,dc in [(1,0),(-1,0),(0,1),(0,-1)]     if 0<=r+dr<9 and 0<=c+dc<9] for r in range(9) for c in range(9)]def groups(b):    seen=set(); out=[]    for p in range(81):        if not b[p] or p in seen: continue        col=int(b[p]); stack=[p]; stones=set([p]); libs=set(); seen.add(p)        while stack:            u=stack.pop()            for v in AD[u]:                if b[v]==0: libs.add(v)                elif b[v]==col and v not in seen:                    seen.add(v); stones.add(v); stack.append(v)        out.append((col,stones,libs))    return outdef show(b):    return "\n".join("".join(".XO"[{0:0,1:1,-1:2}[int(v)]] for v in b[r*9:(r+1)*9]) for r in range(9))open("/work/analysis_helpers.py","w").write('import numpy as np\nAD='+repr(AD)+'\n'+    '''def groups(b):    seen=set(); out=[]    for p in range(81):        if not b[p] or p in seen: continue        col=int(b[p]); stack=[p]; stones={p}; libs=set(); seen.add(p)        while stack:            u=stack.pop()            for v in AD[u]:                if b[v]==0: libs.add(v)                elif b[v]==col and v not in seen:                    seen.add(v); stones.add(v); stack.append(v)        out.append((col,stones,libs))    return outdef show(b):    return "\\n".join("".join({0:".",1:"X",-1:"O"}[int(v)] for v in b[r*9:(r+1)*9]) for r in range(9))''')losses=[(i,g) for i,g in enumerate(games) if g["opponent"]=="strong" and g["result"]<0]for i,g in losses[:5]:    print("\nGAME",i,"net",g["net_color"],"margin",g["margin"],"end",g["end"])    col=1 if g["net_color"]=="black" else -1    events=[]    for t,m in enumerate(g["moves"]):        if t==0 or m["by"]!="opponent": continue        lost=int(np.sum(boards[i,t-1]==col)-np.sum(boards[i,t]==col))        if lost>=4:            events.append((lost,t))    for lost,t in sorted(events,reverse=True)[:2]:        print("LOST",lost,"at",t,"opponent move",g["moves"][t]["move"])        s=max(0,t-3)        print("Board before own move",s+1,"\n"+show(boards[i,s]))        print("Sequence",g["moves"][s+1:t+1])        print("Own groups then",[(len(st),sorted(li)) for c,st,li in groups(boards[i,s]) if c==col])

OUTPUT of Python run 3 of at most 120:


GAME 32 net black margin -8.5 end two passesLOST 4 at 67 opponent move 73Board before own move 65 ......O...X.X.XXXXX.X.X.XOX.X.XOXXOXX.XXOOXOOOXOOOOOOOOOOO.OXXOOXXOXO.OXX..O.O.X.Sequence [{'ply': 65, 'by': 'opponent', 'move': 80}, {'ply': 66, 'by': 'net', 'move': 74, 'top5': [[73, 0.5986], [74, 0.06], [7, 0.0571], [5, 0.0445], [4, 0.0402]]}, {'ply': 67, 'by': 'opponent', 'move': 73}]Own groups then [(1, [1, 9, 11, 19]), (1, [3, 11, 13, 21]), (10, [5, 7, 8, 13, 23]), (1, [9, 19, 27]), (1, [11, 19, 21, 29]), (1, [13, 21, 23]), (1, [19, 27, 29, 37]), (3, [21, 29, 37]), (1, [27, 37]), (1, [37]), (2, [69]), (2, [73, 74]), (1, [58, 76]), (1, [80]), (1, [73]), (1, [78, 80])]
GAME 33 net black margin -24.5 end two passesLOST 14 at 71 opponent move 3Board before own move 69 X.O.XXXO...XOXOXOOX.X.XOXOOXX.OXOXOOX.XXOOXOOXXOXOXXXO.XOXOXOOOXXOOOO.OO.XO..O...Sequence [{'ply': 69, 'by': 'opponent', 'move': 21}, {'ply': 70, 'by': 'net', 'move': 78, 'top5': [[29, 0.3312], [9, 0.2112], [10, 0.2112], [1, 0.0997], [78, 0.0739]]}, {'ply': 71, 'by': 'opponent', 'move': 3}]Own groups then [(1, [1, 9]), (14, [3, 21]), (2, [10, 19, 21, 29]), (10, [9, 19, 29, 37, 54, 72]), (4, [29, 37])]LOST 4 at 59 opponent move 77Board before own move 57 X.O.XXXO...XOXOXO.X.X.XOXOO.X..XOXOO..XXOOXOO.X.XOXXXO..OXOXOOOXX.OOOXOO..O...XX.Sequence [{'ply': 57, 'by': 'opponent', 'move': 65}, {'ply': 58, 'by': 'net', 'move': 80, 'top5': [[80, 0.9866], [77, 0.0019], [76, 0.0016], [55, 0.0013], [36, 0.0012]]}, {'ply': 59, 'by': 'opponent', 'move': 77}]Own groups then [(1, [1, 9]), (14, [3, 21, 30]), (2, [10, 19, 21, 29]), (1, [9, 19, 27]), (1, [19, 27, 29, 37]), (4, [29, 30, 37, 47]), (1, [37, 45, 47, 55]), (2, [54, 55, 65, 72, 73]), (3, [77, 80])]
GAME 34 net black margin -17.5 end two passesLOST 7 at 61 opponent move 7Board before own move 59 X.X.XOX.....OXOXOOX.XXOOXOOX.X.XOXXX.XXXOOOXO...XOOOO...XXXXX.O.OOOOOOX....O.O.X.Sequence [{'ply': 59, 'by': 'opponent', 'move': 53}, {'ply': 60, 'by': 'net', 'move': 46, 'top5': [[61, 0.3527], [46, 0.1666], [55, 0.1298], [10, 0.1117], [11, 0.0712]]}, {'ply': 61, 'by': 'opponent', 'move': 7}]Own groups then [(1, [1, 9]), (1, [1, 3, 11]), (2, [3]), (7, [7]), (2, [9, 19, 28, 36]), (12, [11, 19, 28, 30, 36, 46, 47, 55, 61]), (1, [30]), (2, [61, 71, 78, 80])]
GAME 35 net black margin -26.5 end two passesLOST 8 at 93 opponent move 36Board before own move 91 .O.OOOOO.OOOOOOO.O.OXOXXXOXXOXOX.XXX.XXOOXX.XXOXOOOXX..OXXOXX.XOOOOOOXXX.XXO.OOO.Sequence [{'ply': 91, 'by': 'opponent', 'move': 54}, {'ply': 92, 'by': 'net', 'move': 16, 'top5': [[16, 0.9999], [36, 0.0001], [81, 0.0], [18, 0.0], [80, 0.0]]}, {'ply': 93, 'by': 'opponent', 'move': 36}]Own groups then [(7, [36]), (19, [32, 43, 53, 61, 80]), (1, [18, 36]), (1, [36, 54]), (2, [72])]LOST 7 at 73 opponent move 36Board before own move 71 .O.O.OOO.OOO.OOOX..OXOXXX.XXOXOX.XX..XXOOXX.XXOXOOOXX..OXXOXX.XO.OOOOXXX..XOXX.O.Sequence [{'ply': 71, 'by': 'opponent', 'move': 78}, {'ply': 72, 'by': 'net', 'move': 73, 'top5': [[73, 0.9412], [36, 0.0446], [17, 0.0121], [81, 0.0013], [18, 0.0003]]}, {'ply': 73, 'by': 'opponent', 'move': 36}]Own groups then [(1, [17, 25]), (7, [36]), (16, [25, 32, 35, 43, 53, 61, 78, 80]), (1, [17, 25, 35]), (1, [18, 36]), (1, [35, 43, 53]), (1, [36, 54]), (1, [73]), (2, [78])]
GAME 36 net black margin -19.5 end two passesLOST 7 at 57 opponent move 3Board before own move 55 .OX.XO...X.XXXOO...OOOXOOX..XXOXOX.X..OOOXXX.XXO.X.XX..O.OOXX..X..XOX.O..........Sequence [{'ply': 55, 'by': 'opponent', 'move': 10}, {'ply': 56, 'by': 'net', 'move': 65, 'top5': [[64, 0.1578], [27, 0.091], [65, 0.091], [61, 0.0866], [16, 0.0709]]}, {'ply': 57, 'by': 'opponent', 'move': 3}]Own groups then [(7, [3, 10]), (1, [0, 10, 18]), (1, [16, 26, 34]), (2, [27, 37]), (9, [34, 44, 50, 53, 61, 69, 77]), (1, [26, 34, 44]), (2, [36, 37, 54]), (1, [48, 50]), (1, [54, 64, 72]), (1, [65, 75])]LOST 4 at 91 opponent move 2Board before own move 89 .O.O.O.X.XOXXXOOOOOOOOXOO.O.X.OOOXO.OOOOOXXXO..OXXXXXOOOOOOXXX.XXXXOXOO.X.X.XXO.XSequence [{'ply': 89, 'by': 'opponent', 'move': 4}, {'ply': 90, 'by': 'net', 'move': 71, 'top5': [[71, 0.4575], [75, 0.3938], [29, 0.0971], [81, 0.0278], [79, 0.0088]]}, {'ply': 91, 'by': 'opponent', 'move': 2}]Own groups then [(1, [6, 8]), (1, [0]), (4, [2, 4]), (1, [27, 29]), (15, [62, 75]), (6, [73, 75]), (1, [71, 79])]


GAME 32 net black margin -8.5 end two passesLOST 4 at 67 opponent move 73Board before own move 65 ......O...X.X.XXXXX.X.X.XOX.X.XOXXOXX.XXOOXOOOXOOOOOOOOOOO.OXXOOXXOXO.OXX..O.O.X.Sequence [{'ply': 65, 'by': 'opponent', 'move': 80}, {'ply': 66, 'by': 'net', 'move': 74, 'top5': [[73, 0.5986], [74, 0.06], [7, 0.0571], [5, 0.0445], [4, 0.0402]]}, {'ply': 67, 'by': 'opponent', 'move': 73}]Own groups then [(1, [1, 9, 11, 19]), (1, [3, 11, 13, 21]), (10, [5, 7, 8, 13, 23]), (1, [9, 19, 27]), (1, [11, 19, 21, 29]), (1, [13, 21, 23]), (1, [19, 27, 29, 37]), (3, [21, 29, 37]), (1, [27, 37]), (1, [37]), (2, [69]), (2, [73, 74]), (1, [58, 76]), (1, [80]), (1, [73]), (1, [78, 80])]
GAME 33 net black margin -24.5 end two passesLOST 14 at 71 opponent move 3Board before own move 69 X.O.XXXO...XOXOXOOX.X.XOXOOXX.OXOXOOX.XXOOXOOXXOXOXXXO.XOXOXOOOXXOOOO.OO.XO..O...Sequence [{'ply': 69, 'by': 'opponent', 'move': 21}, {'ply': 70, 'by': 'net', 'move': 78, 'top5': [[29, 0.3312], [9, 0.2112], [10, 0.2112], [1, 0.0997], [78, 0.0739]]}, {'ply': 71, 'by': 'opponent', 'move': 3}]Own groups then [(1, [1, 9]), (14, [3, 21]), (2, [10, 19, 21, 29]), (10, [9, 19, 29, 37, 54, 72]), (4, [29, 37])]LOST 4 at 59 opponent move 77Board before own move 57 X.O.XXXO...XOXOXO.X.X.XOXOO.X..XOXOO..XXOOXOO.X.XOXXXO..OXOXOOOXX.OOOXOO..O...XX.Sequence [{'ply': 57, 'by': 'opponent', 'move': 65}, {'ply': 58, 'by': 'net', 'move': 80, 'top5': [[80, 0.9866], [77, 0.0019], [76, 0.0016], [55, 0.0013], [36, 0.0012]]}, {'ply': 59, 'by': 'opponent', 'move': 77}]Own groups then [(1, [1, 9]), (14, [3, 21, 30]), (2, [10, 19, 21, 29]), (1, [9, 19, 27]), (1, [19, 27, 29, 37]), (4, [29, 30, 37, 47]), (1, [37, 45, 47, 55]), (2, [54, 55, 65, 72, 73]), (3, [77, 80])]
GAME 34 net black margin -17.5 end two passesLOST 7 at 61 opponent move 7Board before own move 59 X.X.XOX.....OXOXOOX.XXOOXOOX.X.XOXXX.XXXOOOXO...XOOOO...XXXXX.O.OOOOOOX....O.O.X.Sequence [{'ply': 59, 'by': 'opponent', 'move': 53}, {'ply': 60, 'by': 'net', 'move': 46, 'top5': [[61, 0.3527], [46, 0.1666], [55, 0.1298], [10, 0.1117], [11, 0.0712]]}, {'ply': 61, 'by': 'opponent', 'move': 7}]Own groups then [(1, [1, 9]), (1, [1, 3, 11]), (2, [3]), (7, [7]), (2, [9, 19, 28, 36]), (12, [11, 19, 28, 30, 36, 46, 47, 55, 61]), (1, [30]), (2, [61, 71, 78, 80])]
GAME 35 net black margin -26.5 end two passesLOST 8 at 93 opponent move 36Board before own move 91 .O.OOOOO.OOOOOOO.O.OXOXXXOXXOXOX.XXX.XXOOXX.XXOXOOOXX..OXXOXX.XOOOOOOXXX.XXO.OOO.Sequence [{'ply': 91, 'by': 'opponent', 'move': 54}, {'ply': 92, 'by': 'net', 'move': 16, 'top5': [[16, 0.9999], [36, 0.0001], [81, 0.0], [18, 0.0], [80, 0.0]]}, {'ply': 93, 'by': 'opponent', 'move': 36}]Own groups then [(7, [36]), (19, [32, 43, 53, 61, 80]), (1, [18, 36]), (1, [36, 54]), (2, [72])]LOST 7 at 73 opponent move 36Board before own move 71 .O.O.OOO.OOO.OOOX..OXOXXX.XXOXOX.XX..XXOOXX.XXOXOOOXX..OXXOXX.XO.OOOOXXX..XOXX.O.Sequence [{'ply': 71, 'by': 'opponent', 'move': 78}, {'ply': 72, 'by': 'net', 'move': 73, 'top5': [[73, 0.9412], [36, 0.0446], [17, 0.0121], [81, 0.0013], [18, 0.0003]]}, {'ply': 73, 'by': 'opponent', 'move': 36}]Own groups then [(1, [17, 25]), (7, [36]), (16, [25, 32, 35, 43, 53, 61, 78, 80]), (1, [17, 25, 35]), (1, [18, 36]), (1, [35, 43, 53]), (1, [36, 54]), (1, [73]), (2, [78])]
GAME 36 net black margin -19.5 end two passesLOST 7 at 57 opponent move 3Board before own move 55 .OX.XO...X.XXXOO...OOOXOOX..XXOXOX.X..OOOXXX.XXO.X.XX..O.OOXX..X..XOX.O..........Sequence [{'ply': 55, 'by': 'opponent', 'move': 10}, {'ply': 56, 'by': 'net', 'move': 65, 'top5': [[64, 0.1578], [27, 0.091], [65, 0.091], [61, 0.0866], [16, 0.0709]]}, {'ply': 57, 'by': 'opponent', 'move': 3}]Own groups then [(7, [3, 10]), (1, [0, 10, 18]), (1, [16, 26, 34]), (2, [27, 37]), (9, [34, 44, 50, 53, 61, 69, 77]), (1, [26, 34, 44]), (2, [36, 37, 54]), (1, [48, 50]), (1, [54, 64, 72]), (1, [65, 75])]LOST 4 at 91 opponent move 2Board before own move 89 .O.O.O.X.XOXXXOOOOOOOOXOO.O.X.OOOXO.OOOOOXXXO..OXXXXXOOOOOOXXX.XXXXOXOO.X.X.XXO.XSequence [{'ply': 89, 'by': 'opponent', 'move': 4}, {'ply': 90, 'by': 'net', 'move': 71, 'top5': [[71, 0.4575], [75, 0.3938], [29, 0.0971], [81, 0.0278], [79, 0.0088]]}, {'ply': 91, 'by': 'opponent', 'move': 2}]Own groups then [(1, [6, 8]), (1, [0]), (4, [2, 4]), (1, [27, 29]), (15, [62, 75]), (6, [73, 75]), (1, [71, 79])]

python

import shutilshutil.copyfile("/work/design.py","/work/design_v1.py")s=open("/work/design.py").read()start=s.index("# Second layer")end=s.index("    dest = unit(4,p)",start)s=s[:start]+r'''# Conservative safety seeds: never mark a genuinely atari group safe.# Channels 2/4 are binary danger; 3/5 detect adjacent liberty-bearing# stones, whose liberties must be distinct on the square grid.for p in range(81):    carry(1,0,p)    carry(1,1,p)    for color,dch,zch in [(0,2,3),(1,4,5)]:        dest=unit(dch,p)        add(1,dest,unit(dch,p),1)        add(1,dest,unit(zch,p),-1)        for q in adj[p]:            add(1,dest,unit(color,q),-1)            add(1,dest,unit(dch,q),1)            add(1,dest,unit(zch,q),-1)        dest=unit(zch,p)        B[1][dest]=-4        add(1,dest,unit(dch,p),4)        add(1,dest,unit(zch,p),-8)        for q in adj[p]:            add(1,dest,unit(dch,q),1)            add(1,dest,unit(zch,q),-2)
# Safety propagates through each connected string.for p in range(81):    carry(2,0,p)    carry(2,1,p)    for color,dch,zch,outch in [(0,2,3,2),(1,4,5,3)]:        dest=unit(outch,p)        add(2,dest,unit(dch,p),1)        add(2,dest,unit(zch,p),-1)        for q in adj[p]:            add(2,dest,unit(color,q),-1)            add(2,dest,unit(dch,q),1)            add(2,dest,unit(zch,q),-1)''' + s[end:]open("/work/design.py","w").write(s)overrides={}exec(s)print("Corrected conservative group-safety logic; other coefficients unchanged.")

import shutilshutil.copyfile("/work/design.py","/work/design_v1.py")s=open("/work/design.py").read()start=s.index("# Second layer")end=s.index("    dest = unit(4,p)",start)s=s[:start]+r'''# Conservative safety seeds: never mark a genuinely atari group safe.# Channels 2/4 are binary danger; 3/5 detect adjacent liberty-bearing# stones, whose liberties must be distinct on the square grid.for p in range(81):    carry(1,0,p)    carry(1,1,p)    for color,dch,zch in [(0,2,3),(1,4,5)]:        dest=unit(dch,p)        add(1,dest,unit(dch,p),1)        add(1,dest,unit(zch,p),-1)        for q in adj[p]:            add(1,dest,unit(color,q),-1)            add(1,dest,unit(dch,q),1)            add(1,dest,unit(zch,q),-1)        dest=unit(zch,p)        B[1][dest]=-4        add(1,dest,unit(dch,p),4)        add(1,dest,unit(zch,p),-8)        for q in adj[p]:            add(1,dest,unit(dch,q),1)            add(1,dest,unit(zch,q),-2)
# Safety propagates through each connected string.for p in range(81):    carry(2,0,p)    carry(2,1,p)    for color,dch,zch,outch in [(0,2,3,2),(1,4,5,3)]:        dest=unit(outch,p)        add(2,dest,unit(dch,p),1)        add(2,dest,unit(zch,p),-1)        for q in adj[p]:            add(2,dest,unit(color,q),-1)            add(2,dest,unit(dch,q),1)            add(2,dest,unit(zch,q),-1)''' + s[end:]open("/work/design.py","w").write(s)overrides={}exec(s)print("Corrected conservative group-safety logic; other coefficients unchanged.")

OUTPUT of Python run 4 of at most 120:

Games pass 2 of 25 done.

{"ok": true, "pass": 2, "kind": "games", "per_opponent": {"strong": {"games": 80, "score": 0.5249999761581421, "wins": 42, "losses": 38, "score_as_black": 0.42500001192092896, "score_as_white": 0.625, "mean_margin": 28.325000762939453, "mean_length": 111.85, "stones_captured_per_game": 47.5625, "stones_lost_per_game": 17.75, "passes_per_game": 2.475, "games_with_a_pass_before_move_60": 2, "end_two_passes": 68, "end_repetition": 8, "end_move_limit": 4}, "weak": {"games": 48, "score": 0.8125, "wins": 39, "losses": 9, "score_as_black": 0.75, "score_as_white": 0.875, "mean_margin": 65.5625, "mean_length": 118.54166666666667, "stones_captured_per_game": 70.29166666666667, "stones_lost_per_game": 4.895833333333333, "passes_per_game": 1.8958333333333333, "games_with_a_pass_before_move_60": 2, "end_two_passes": 43, "end_repetition": 5, "end_move_limit": 0}}, "files": "/work/pass_2/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 2, "kind": "games", "per_opponent": {"strong": {"games": 80, "score": 0.5249999761581421, "wins": 42, "losses": 38, "score_as_black": 0.42500001192092896, "score_as_white": 0.625, "mean_margin": 28.325000762939453, "mean_length": 111.85, "stones_captured_per_game": 47.5625, "stones_lost_per_game": 17.75, "passes_per_game": 2.475, "games_with_a_pass_before_move_60": 2, "end_two_passes": 68, "end_repetition": 8, "end_move_limit": 4}, "weak": {"games": 48, "score": 0.8125, "wins": 39, "losses": 9, "score_as_black": 0.75, "score_as_white": 0.875, "mean_margin": 65.5625, "mean_length": 118.54166666666667, "stones_captured_per_game": 70.29166666666667, "stones_lost_per_game": 4.895833333333333, "passes_per_game": 1.8958333333333333, "games_with_a_pass_before_move_60": 2, "end_two_passes": 43, "end_repetition": 5, "end_move_limit": 0}}, "files": "/work/pass_2/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 23.

python

exec(open("/work/analysis_helpers.py").read())games=json.load(open("/work/pass_2/games.json"))boards=np.load("/work/pass_2/boards.npy")stats={}examples=[]for i,g in enumerate(games):    col=1 if g["net_color"]=="black" else -1    for t,m in enumerate(g["moves"]):        if m["by"]!="net": continue        b=np.zeros(81,dtype=int) if t==0 else boards[i,t-1]        gs=groups(b)        endangered=[(len(st),next(iter(li))) for c,st,li in gs if c==col and len(li)==1]        caps=[(len(st),next(iter(li))) for c,st,li in gs if c==-col and len(li)==1]        if endangered and g["opponent"]=="strong":            moved=m["move"]; biggest=max(endangered)            if biggest[0]>=4 and moved!=biggest[1]:                after=groups(boards[i,t])                still=[(len(st),list(li)) for c,st,li in after if c==col and len(li)==1 and biggest[1] in li]                if still: examples.append((biggest[0],i,t,biggest,m["move"],m["top5"]))print("Missed large ataris",len(examples))for size,i,t,big,move,top in sorted(examples,reverse=True)[:10]:    g=games[i]    print("\nGame",i,g["net_color"],g["result"],"ply",t,"atari",big,"move",move,"top",top)    print(show(boards[i,t-1]))    print("groups",[(c,len(st),sorted(li)) for c,st,li in groups(boards[i,t-1]) if len(li)<=2])print("\nLoss endings:")from collections import Counterprint(Counter((g["opponent"],g["end"]) for g in games if g["result"]<0))

exec(open("/work/analysis_helpers.py").read())games=json.load(open("/work/pass_2/games.json"))boards=np.load("/work/pass_2/boards.npy")stats={}examples=[]for i,g in enumerate(games):    col=1 if g["net_color"]=="black" else -1    for t,m in enumerate(g["moves"]):        if m["by"]!="net": continue        b=np.zeros(81,dtype=int) if t==0 else boards[i,t-1]        gs=groups(b)        endangered=[(len(st),next(iter(li))) for c,st,li in gs if c==col and len(li)==1]        caps=[(len(st),next(iter(li))) for c,st,li in gs if c==-col and len(li)==1]        if endangered and g["opponent"]=="strong":            moved=m["move"]; biggest=max(endangered)            if biggest[0]>=4 and moved!=biggest[1]:                after=groups(boards[i,t])                still=[(len(st),list(li)) for c,st,li in after if c==col and len(li)==1 and biggest[1] in li]                if still: examples.append((biggest[0],i,t,biggest,m["move"],m["top5"]))print("Missed large ataris",len(examples))for size,i,t,big,move,top in sorted(examples,reverse=True)[:10]:    g=games[i]    print("\nGame",i,g["net_color"],g["result"],"ply",t,"atari",big,"move",move,"top",top)    print(show(boards[i,t-1]))    print("groups",[(c,len(st),sorted(li)) for c,st,li in groups(boards[i,t-1]) if len(li)<=2])print("\nLoss endings:")from collections import Counterprint(Counter((g["opponent"],g["end"]) for g in games if g["result"]<0))

OUTPUT of Python run 5 of at most 120:

Missed large ataris 274
Game 58 white -1 ply 79 atari (15, 77) move 78 top [[78, 0.6875], [80, 0.3089], [14, 0.0019], [36, 0.0017], [81, 0.0]].OXXX.X.XOXXX..XX.OOOXXXOX.OXXXXXOXX.OOXXOOOOO.OXXOOOO.OOXOOOXOOXXXXOXXX.X.XX..O.groups [(-1, 1, [0]), (1, 1, [7, 17]), (-1, 5, [0, 36]), (-1, 15, [77]), (1, 4, [78, 80]), (-1, 1, [54, 72]), (-1, 1, [78, 80])]
Game 44 white -1 ply 77 atari (13, 78) move 6 top [[6, 0.9893], [45, 0.0023], [36, 0.0022], [79, 0.0021], [54, 0.001]].OO.OO.XOOOOOXOOOOO.OOXXOXOOOXOXXXXX.XXXXOOXO.XXXOOOOO.OOXXXOXX..XXXOOX..O.XOO...groups [(-1, 9, [3, 6]), (1, 1, [6]), (-1, 13, [78]), (-1, 2, [54, 64]), (1, 3, [71, 79])]
Game 44 white -1 ply 75 atari (13, 78) move 8 top [[8, 0.9999], [3, 0.0], [45, 0.0], [36, 0.0], [79, 0.0]].OO.OOXX.OOOOXOOOOO.OOXXOXOOOXOXXXXX.XXXXOOXO.XXXOOOOO.OOXXXOXX..XXXOOX..O.XOO...groups [(-1, 8, [3, 8]), (1, 2, [8]), (-1, 13, [78]), (-1, 2, [54, 64]), (1, 3, [71, 79])]
Game 44 white -1 ply 73 atari (13, 78) move 2 top [[2, 0.9825], [45, 0.0038], [36, 0.0036], [79, 0.0035], [54, 0.0016]].O.XOO.X.OOOOXOOOOO.OOXXOXOOOXOXXXXX.XXXXOOXO.XXXOOOOO.OOXXXOXX..XXXOOX..O.XOO...groups [(1, 1, [2]), (-1, 8, [6, 8]), (1, 1, [6, 8]), (-1, 13, [78]), (-1, 2, [54, 64]), (1, 3, [71, 79])]
Game 44 white -1 ply 71 atari (13, 78) move 5 top [[5, 0.9999], [2, 0.0001], [45, 0.0], [36, 0.0], [79, 0.0]].O.XO.XXXOOOOXOOOOO.OOXXOXOOOXOXXXXX.XXXXOOXO.XXXOOOOO.OOXXXOXX..XXXOOX..O.XOO...groups [(1, 1, [2]), (-1, 1, [5]), (1, 3, [5]), (-1, 6, [5]), (-1, 13, [78]), (-1, 2, [54, 64]), (1, 3, [71, 79])]
Game 37 black -1 ply 96 atari (11, 0) move 68 top [[68, 0.7588], [81, 0.2403], [64, 0.0008], [73, 0.0001], [53, 0.0]].XXXXOOO.OXXXO.OOOOOXOOOOXXOOXOOOOX.OOXOOXX.XXOXOXXXX..OOOOXX.XO.OOX.XX.O.OOOXX.Xgroups [(1, 11, [0]), (1, 3, [35, 43]), (1, 1, [54]), (1, 1, [68]), (1, 1, [71, 79])]
Game 37 black -1 ply 94 atari (11, 0) move 50 top [[50, 0.5512], [68, 0.2477], [81, 0.1502], [36, 0.0352], [54, 0.0151]].XXXXOOO.OXXXO.OOOOOXOOOOXXOOXOOOOX..OXOOXX.XXOXOX.XX..OOOOXX.XO.OOX.XX.O.OOOXX.Xgroups [(1, 11, [0]), (1, 3, [35, 43]), (1, 1, [36, 54]), (1, 1, [50]), (1, 1, [68]), (1, 1, [71, 79])]
Game 37 black -1 ply 92 atari (11, 0) move 45 top [[64, 0.5271], [45, 0.2038], [54, 0.1119], [63, 0.0788], [50, 0.0455]].XXXXOOO.OXXXO.OOOOOXOOOOXXOOXOOOOX..OXOOXX.X.OXOX.XX..OOOOXX.X..OOX.XX.O.OOOXX.Xgroups [(1, 11, [0]), (1, 3, [35, 43]), (1, 1, [50]), (1, 1, [68]), (-1, 1, [63, 73]), (1, 1, [71, 79])]
Game 37 black -1 ply 90 atari (11, 0) move 49 top [[49, 1.0], [64, 0.0], [45, 0.0], [54, 0.0], [73, 0.0]].XXXXOOO.OXXXO.OOOOOXOOOOXXOOXOOOOX..OXOOXX.X.OXO.OXX..OOOOXX.X..OOX.XX.O..OOXX.Xgroups [(1, 11, [0]), (1, 3, [35, 43]), (-1, 1, [49]), (1, 1, [68]), (-1, 1, [63, 73]), (1, 1, [71, 79])]
Game 37 black -1 ply 88 atari (11, 0) move 26 top [[26, 0.7542], [35, 0.1523], [68, 0.056], [81, 0.0375], [53, 0.0]].XXXXOOO.OXXXO.OOOOOXOOOOX.OOXOOOOX.XOXOOXX.XXOXO.OXX.XOOOOXX.XXXOOX.XX..XXOOXX.Xgroups [(1, 11, [0]), (1, 7, [72]), (-1, 1, [49]), (1, 1, [68]), (1, 1, [71, 79])]
Loss endings:Counter({('strong', 'two passes'): 33, ('weak', 'two passes'): 6, ('strong', 'repetition'): 5, ('weak', 'repetition'): 3})

Missed large ataris 274
Game 58 white -1 ply 79 atari (15, 77) move 78 top [[78, 0.6875], [80, 0.3089], [14, 0.0019], [36, 0.0017], [81, 0.0]].OXXX.X.XOXXX..XX.OOOXXXOX.OXXXXXOXX.OOXXOOOOO.OXXOOOO.OOXOOOXOOXXXXOXXX.X.XX..O.groups [(-1, 1, [0]), (1, 1, [7, 17]), (-1, 5, [0, 36]), (-1, 15, [77]), (1, 4, [78, 80]), (-1, 1, [54, 72]), (-1, 1, [78, 80])]
Game 44 white -1 ply 77 atari (13, 78) move 6 top [[6, 0.9893], [45, 0.0023], [36, 0.0022], [79, 0.0021], [54, 0.001]].OO.OO.XOOOOOXOOOOO.OOXXOXOOOXOXXXXX.XXXXOOXO.XXXOOOOO.OOXXXOXX..XXXOOX..O.XOO...groups [(-1, 9, [3, 6]), (1, 1, [6]), (-1, 13, [78]), (-1, 2, [54, 64]), (1, 3, [71, 79])]
Game 44 white -1 ply 75 atari (13, 78) move 8 top [[8, 0.9999], [3, 0.0], [45, 0.0], [36, 0.0], [79, 0.0]].OO.OOXX.OOOOXOOOOO.OOXXOXOOOXOXXXXX.XXXXOOXO.XXXOOOOO.OOXXXOXX..XXXOOX..O.XOO...groups [(-1, 8, [3, 8]), (1, 2, [8]), (-1, 13, [78]), (-1, 2, [54, 64]), (1, 3, [71, 79])]
Game 44 white -1 ply 73 atari (13, 78) move 2 top [[2, 0.9825], [45, 0.0038], [36, 0.0036], [79, 0.0035], [54, 0.0016]].O.XOO.X.OOOOXOOOOO.OOXXOXOOOXOXXXXX.XXXXOOXO.XXXOOOOO.OOXXXOXX..XXXOOX..O.XOO...groups [(1, 1, [2]), (-1, 8, [6, 8]), (1, 1, [6, 8]), (-1, 13, [78]), (-1, 2, [54, 64]), (1, 3, [71, 79])]
Game 44 white -1 ply 71 atari (13, 78) move 5 top [[5, 0.9999], [2, 0.0001], [45, 0.0], [36, 0.0], [79, 0.0]].O.XO.XXXOOOOXOOOOO.OOXXOXOOOXOXXXXX.XXXXOOXO.XXXOOOOO.OOXXXOXX..XXXOOX..O.XOO...groups [(1, 1, [2]), (-1, 1, [5]), (1, 3, [5]), (-1, 6, [5]), (-1, 13, [78]), (-1, 2, [54, 64]), (1, 3, [71, 79])]
Game 37 black -1 ply 96 atari (11, 0) move 68 top [[68, 0.7588], [81, 0.2403], [64, 0.0008], [73, 0.0001], [53, 0.0]].XXXXOOO.OXXXO.OOOOOXOOOOXXOOXOOOOX.OOXOOXX.XXOXOXXXX..OOOOXX.XO.OOX.XX.O.OOOXX.Xgroups [(1, 11, [0]), (1, 3, [35, 43]), (1, 1, [54]), (1, 1, [68]), (1, 1, [71, 79])]
Game 37 black -1 ply 94 atari (11, 0) move 50 top [[50, 0.5512], [68, 0.2477], [81, 0.1502], [36, 0.0352], [54, 0.0151]].XXXXOOO.OXXXO.OOOOOXOOOOXXOOXOOOOX..OXOOXX.XXOXOX.XX..OOOOXX.XO.OOX.XX.O.OOOXX.Xgroups [(1, 11, [0]), (1, 3, [35, 43]), (1, 1, [36, 54]), (1, 1, [50]), (1, 1, [68]), (1, 1, [71, 79])]
Game 37 black -1 ply 92 atari (11, 0) move 45 top [[64, 0.5271], [45, 0.2038], [54, 0.1119], [63, 0.0788], [50, 0.0455]].XXXXOOO.OXXXO.OOOOOXOOOOXXOOXOOOOX..OXOOXX.X.OXOX.XX..OOOOXX.X..OOX.XX.O.OOOXX.Xgroups [(1, 11, [0]), (1, 3, [35, 43]), (1, 1, [50]), (1, 1, [68]), (-1, 1, [63, 73]), (1, 1, [71, 79])]
Game 37 black -1 ply 90 atari (11, 0) move 49 top [[49, 1.0], [64, 0.0], [45, 0.0], [54, 0.0], [73, 0.0]].XXXXOOO.OXXXO.OOOOOXOOOOXXOOXOOOOX..OXOOXX.X.OXO.OXX..OOOOXX.X..OOX.XX.O..OOXX.Xgroups [(1, 11, [0]), (1, 3, [35, 43]), (-1, 1, [49]), (1, 1, [68]), (-1, 1, [63, 73]), (1, 1, [71, 79])]
Game 37 black -1 ply 88 atari (11, 0) move 26 top [[26, 0.7542], [35, 0.1523], [68, 0.056], [81, 0.0375], [53, 0.0]].XXXXOOO.OXXXO.OOOOOXOOOOX.OOXOOOOX.XOXOOXX.XXOXO.OXX.XOOOOXX.XXXOOX.XX..XXOOXX.Xgroups [(1, 11, [0]), (1, 7, [72]), (-1, 1, [49]), (1, 1, [68]), (1, 1, [71, 79])]
Loss endings:Counter({('strong', 'two passes'): 33, ('weak', 'two passes'): 6, ('strong', 'repetition'): 5, ('weak', 'repetition'): 3})

python

import shutilshutil.copyfile("/work/design.py","/work/design_v2.py")s=open("/work/design.py").read()s=s.replace('seed_shared_liberty=1.0,','seed_shared_liberty=1.0, positional_scale=2.0, low_liberty=1.5,')s=s.replace('    add(3,unit(5,p),unit(4,p),1)',r'''    # Prefer moves with room to grow, not just moves that evade atari.    dest=unit(5,p)    B[3][dest]=3-len(adj[p])    for q in adj[p]:        add(3,dest,unit(0,q),-1)        add(3,dest,unit(1,q),1)        add(3,dest,unit(2,q),2)        add(3,dest,unit(3,q),-3)''')s=s.replace('add(5,dest,unit(5,p),-1.5)','add(5,dest,unit(5,p),-P["low_liberty"])')s=s.replace('OB[p] = -20 + positional','OB[p] = -20 + P["positional_scale"]*positional')open("/work/design.py","w").write(s)overrides=dict(friend_adj=1.4,enemy_adj=0.5,friend_diag=0.6,               friend_jump=0.25,friend_knight=0.1,               crowd=2.0,rescue=6.0,self_atari=9.0)open("/work/params_v3.json","w").write(json.dumps(overrides))exec(s)print(P)

import shutilshutil.copyfile("/work/design.py","/work/design_v2.py")s=open("/work/design.py").read()s=s.replace('seed_shared_liberty=1.0,','seed_shared_liberty=1.0, positional_scale=2.0, low_liberty=1.5,')s=s.replace('    add(3,unit(5,p),unit(4,p),1)',r'''    # Prefer moves with room to grow, not just moves that evade atari.    dest=unit(5,p)    B[3][dest]=3-len(adj[p])    for q in adj[p]:        add(3,dest,unit(0,q),-1)        add(3,dest,unit(1,q),1)        add(3,dest,unit(2,q),2)        add(3,dest,unit(3,q),-3)''')s=s.replace('add(5,dest,unit(5,p),-1.5)','add(5,dest,unit(5,p),-P["low_liberty"])')s=s.replace('OB[p] = -20 + positional','OB[p] = -20 + P["positional_scale"]*positional')open("/work/design.py","w").write(s)overrides=dict(friend_adj=1.4,enemy_adj=0.5,friend_diag=0.6,               friend_jump=0.25,friend_knight=0.1,               crowd=2.0,rescue=6.0,self_atari=9.0)open("/work/params_v3.json","w").write(json.dumps(overrides))exec(s)print(P)

OUTPUT of Python run 6 of at most 120:

{'capture': 7.0, 'rescue': 6.0, 'self_atari': 9.0, 'eye': 11.0, 'crowd': 2.0, 'pass_bias': -2.0, 'friend_adj': 1.4, 'enemy_adj': 0.5, 'friend_diag': 0.6, 'enemy_diag': 0.1, 'friend_jump': 0.25, 'enemy_jump': 0.15, 'friend_knight': 0.1, 'enemy_knight': 0.05, 'seed_shared_liberty': 1.0, 'positional_scale': 2.0, 'low_liberty': 1.5}

{'capture': 7.0, 'rescue': 6.0, 'self_atari': 9.0, 'eye': 11.0, 'crowd': 2.0, 'pass_bias': -2.0, 'friend_adj': 1.4, 'enemy_adj': 0.5, 'friend_diag': 0.6, 'enemy_diag': 0.1, 'friend_jump': 0.25, 'enemy_jump': 0.15, 'friend_knight': 0.1, 'enemy_knight': 0.05, 'seed_shared_liberty': 1.0, 'positional_scale': 2.0, 'low_liberty': 1.5}

Games pass 3 of 25 done.

{"ok": true, "pass": 3, "kind": "games", "per_opponent": {"strong": {"games": 80, "score": 0.48750001192092896, "wins": 39, "losses": 41, "score_as_black": 0.4749999940395355, "score_as_white": 0.5, "mean_margin": 25.037500381469727, "mean_length": 111.5, "stones_captured_per_game": 49.85, "stones_lost_per_game": 15.5875, "passes_per_game": 2.8375, "games_with_a_pass_before_move_60": 4, "end_two_passes": 78, "end_repetition": 2, "end_move_limit": 0}, "weak": {"games": 48, "score": 0.8541666865348816, "wins": 41, "losses": 7, "score_as_black": 0.7083333134651184, "score_as_white": 1.0, "mean_margin": 63.89583206176758, "mean_length": 123.22916666666667, "stones_captured_per_game": 83.54166666666667, "stones_lost_per_game": 4.416666666666667, "passes_per_game": 2.7083333333333335, "games_with_a_pass_before_move_60": 1, "end_two_passes": 46, "end_repetition": 1, "end_move_limit": 1}}, "files": "/work/pass_3/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 3, "kind": "games", "per_opponent": {"strong": {"games": 80, "score": 0.48750001192092896, "wins": 39, "losses": 41, "score_as_black": 0.4749999940395355, "score_as_white": 0.5, "mean_margin": 25.037500381469727, "mean_length": 111.5, "stones_captured_per_game": 49.85, "stones_lost_per_game": 15.5875, "passes_per_game": 2.8375, "games_with_a_pass_before_move_60": 4, "end_two_passes": 78, "end_repetition": 2, "end_move_limit": 0}, "weak": {"games": 48, "score": 0.8541666865348816, "wins": 41, "losses": 7, "score_as_black": 0.7083333134651184, "score_as_white": 1.0, "mean_margin": 63.89583206176758, "mean_length": 123.22916666666667, "stones_captured_per_game": 83.54166666666667, "stones_lost_per_game": 4.416666666666667, "passes_per_game": 2.7083333333333335, "games_with_a_pass_before_move_60": 1, "end_two_passes": 46, "end_repetition": 1, "end_move_limit": 1}}, "files": "/work/pass_3/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 22.

python

import shutil, jsonshutil.copyfile("/work/design.py","/work/design_v3.py")s=open("/work/design.py").read()s=s.replace('low_liberty=1.5,','low_liberty=1.5, weak_friend=1.0, weak_enemy=2.0,')a=s.index('    dest = unit(4,p)',s.index('# Safety propagates'))b=s.index('# Fourth layer',a)s=s[:a]+r'''    # Local liberty pressure, reduced for already connected stones.    for color,ch in [(0,4),(1,5)]:        dest=unit(ch,p)        add(2,dest,unit(color,p),3)        B[2][dest]=-len(adj[p])        for q in adj[p]:            add(2,dest,unit(0,q),1)            add(2,dest,unit(1,q),1)            add(2,dest,unit(color,q),-0.5)
''' + s[b:]a=s.index('    add(3,unit(4,p),unit(4,p),1)')b=s.index('    # Prefer moves with room',a)s=s[:a]+r'''    # Urgency to support pressured stones and reduce enemy liberties.    for q in adj[p]:        add(3,unit(4,p),unit(4,q),P["weak_friend"])        add(3,unit(4,p),unit(5,q),P["weak_enemy"])''' + s[b:]s=s.replace('    add(5,unit(3,p),unit(4,p),1)',r'''    dest=unit(3,p)    B[5][dest]=1-len(adj[p])    for q in adj[p]:        add(5,dest,unit(0,q),1)    for q in diag[p]:        add(5,dest,unit(1,q),-0.5 if len(adj[p])==4 else -1)''')s=s.replace('add(5,dest,unit(5,p),-P["low_liberty"])','''add(5,dest,unit(5,p),-P["low_liberty"])    add(5,dest,unit(4,p),1)''')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v3.json"))exec(s)print("Added proactive local-liberty pressure and support.")print(P)

import shutil, jsonshutil.copyfile("/work/design.py","/work/design_v3.py")s=open("/work/design.py").read()s=s.replace('low_liberty=1.5,','low_liberty=1.5, weak_friend=1.0, weak_enemy=2.0,')a=s.index('    dest = unit(4,p)',s.index('# Safety propagates'))b=s.index('# Fourth layer',a)s=s[:a]+r'''    # Local liberty pressure, reduced for already connected stones.    for color,ch in [(0,4),(1,5)]:        dest=unit(ch,p)        add(2,dest,unit(color,p),3)        B[2][dest]=-len(adj[p])        for q in adj[p]:            add(2,dest,unit(0,q),1)            add(2,dest,unit(1,q),1)            add(2,dest,unit(color,q),-0.5)
''' + s[b:]a=s.index('    add(3,unit(4,p),unit(4,p),1)')b=s.index('    # Prefer moves with room',a)s=s[:a]+r'''    # Urgency to support pressured stones and reduce enemy liberties.    for q in adj[p]:        add(3,unit(4,p),unit(4,q),P["weak_friend"])        add(3,unit(4,p),unit(5,q),P["weak_enemy"])''' + s[b:]s=s.replace('    add(5,unit(3,p),unit(4,p),1)',r'''    dest=unit(3,p)    B[5][dest]=1-len(adj[p])    for q in adj[p]:        add(5,dest,unit(0,q),1)    for q in diag[p]:        add(5,dest,unit(1,q),-0.5 if len(adj[p])==4 else -1)''')s=s.replace('add(5,dest,unit(5,p),-P["low_liberty"])','''add(5,dest,unit(5,p),-P["low_liberty"])    add(5,dest,unit(4,p),1)''')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v3.json"))exec(s)print("Added proactive local-liberty pressure and support.")print(P)

OUTPUT of Python run 7 of at most 120:

Added proactive local-liberty pressure and support.{'capture': 7.0, 'rescue': 6.0, 'self_atari': 9.0, 'eye': 11.0, 'crowd': 2.0, 'pass_bias': -2.0, 'friend_adj': 1.4, 'enemy_adj': 0.5, 'friend_diag': 0.6, 'enemy_diag': 0.1, 'friend_jump': 0.25, 'enemy_jump': 0.15, 'friend_knight': 0.1, 'enemy_knight': 0.05, 'seed_shared_liberty': 1.0, 'positional_scale': 2.0, 'low_liberty': 1.5, 'weak_friend': 1.0, 'weak_enemy': 2.0}

Added proactive local-liberty pressure and support.{'capture': 7.0, 'rescue': 6.0, 'self_atari': 9.0, 'eye': 11.0, 'crowd': 2.0, 'pass_bias': -2.0, 'friend_adj': 1.4, 'enemy_adj': 0.5, 'friend_diag': 0.6, 'enemy_diag': 0.1, 'friend_jump': 0.25, 'enemy_jump': 0.15, 'friend_knight': 0.1, 'enemy_knight': 0.05, 'seed_shared_liberty': 1.0, 'positional_scale': 2.0, 'low_liberty': 1.5, 'weak_friend': 1.0, 'weak_enemy': 2.0}

Games pass 4 of 25 done.

{"ok": true, "pass": 4, "kind": "games", "per_opponent": {"strong": {"games": 80, "score": 0.550000011920929, "wins": 44, "losses": 36, "score_as_black": 0.4749999940395355, "score_as_white": 0.625, "mean_margin": 35.45000076293945, "mean_length": 115.8375, "stones_captured_per_game": 54.8, "stones_lost_per_game": 15.65, "passes_per_game": 1.9125, "games_with_a_pass_before_move_60": 1, "end_two_passes": 79, "end_repetition": 1, "end_move_limit": 0}, "weak": {"games": 48, "score": 0.8541666865348816, "wins": 41, "losses": 7, "score_as_black": 0.7083333134651184, "score_as_white": 1.0, "mean_margin": 59.625, "mean_length": 121.33333333333333, "stones_captured_per_game": 76.25, "stones_lost_per_game": 5.208333333333333, "passes_per_game": 2.9791666666666665, "games_with_a_pass_before_move_60": 1, "end_two_passes": 47, "end_repetition": 1, "end_move_limit": 0}}, "files": "/work/pass_4/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 4, "kind": "games", "per_opponent": {"strong": {"games": 80, "score": 0.550000011920929, "wins": 44, "losses": 36, "score_as_black": 0.4749999940395355, "score_as_white": 0.625, "mean_margin": 35.45000076293945, "mean_length": 115.8375, "stones_captured_per_game": 54.8, "stones_lost_per_game": 15.65, "passes_per_game": 1.9125, "games_with_a_pass_before_move_60": 1, "end_two_passes": 79, "end_repetition": 1, "end_move_limit": 0}, "weak": {"games": 48, "score": 0.8541666865348816, "wins": 41, "losses": 7, "score_as_black": 0.7083333134651184, "score_as_white": 1.0, "mean_margin": 59.625, "mean_length": 121.33333333333333, "stones_captured_per_game": 76.25, "stones_lost_per_game": 5.208333333333333, "passes_per_game": 2.9791666666666665, "games_with_a_pass_before_move_60": 1, "end_two_passes": 47, "end_repetition": 1, "end_move_limit": 0}}, "files": "/work/pass_4/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 21.

python

import shutilshutil.copyfile("/work/design.py","/work/design_v4.py")s=open("/work/design.py").read()s=s.replace('weak_friend=1.0, weak_enemy=2.0,','weak_friend=1.0, weak_enemy=2.0, decisiveness=1.6,')s=s.replace('for k in range(6):\n    save(', 'OW *= P["decisiveness"]\nOB *= P["decisiveness"]\n\nfor k in range(6):\n    save(')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v3.json"))exec(s)print("Increased policy decisiveness, with unchanged move ranking.")

import shutilshutil.copyfile("/work/design.py","/work/design_v4.py")s=open("/work/design.py").read()s=s.replace('weak_friend=1.0, weak_enemy=2.0,','weak_friend=1.0, weak_enemy=2.0, decisiveness=1.6,')s=s.replace('for k in range(6):\n    save(', 'OW *= P["decisiveness"]\nOB *= P["decisiveness"]\n\nfor k in range(6):\n    save(')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v3.json"))exec(s)print("Increased policy decisiveness, with unchanged move ranking.")

OUTPUT of Python run 8 of at most 120:

Games pass 5 of 25 done.

{"ok": true, "pass": 5, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.5833333134651184, "wins": 56, "losses": 40, "score_as_black": 0.4375, "score_as_white": 0.7291666865348816, "mean_margin": 39.34375, "mean_length": 115.57291666666667, "stones_captured_per_game": 61.333333333333336, "stones_lost_per_game": 15.479166666666666, "passes_per_game": 1.6979166666666667, "games_with_a_pass_before_move_60": 0, "end_two_passes": 89, "end_repetition": 4, "end_move_limit": 3}, "weak": {"games": 32, "score": 0.9375, "wins": 30, "losses": 2, "score_as_black": 0.875, "score_as_white": 1.0, "mean_margin": 69.71875, "mean_length": 123.1875, "stones_captured_per_game": 87.6875, "stones_lost_per_game": 3.125, "passes_per_game": 2.375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 32, "end_repetition": 0, "end_move_limit": 0}}, "files": "/work/pass_5/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 5, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.5833333134651184, "wins": 56, "losses": 40, "score_as_black": 0.4375, "score_as_white": 0.7291666865348816, "mean_margin": 39.34375, "mean_length": 115.57291666666667, "stones_captured_per_game": 61.333333333333336, "stones_lost_per_game": 15.479166666666666, "passes_per_game": 1.6979166666666667, "games_with_a_pass_before_move_60": 0, "end_two_passes": 89, "end_repetition": 4, "end_move_limit": 3}, "weak": {"games": 32, "score": 0.9375, "wins": 30, "losses": 2, "score_as_black": 0.875, "score_as_white": 1.0, "mean_margin": 69.71875, "mean_length": 123.1875, "stones_captured_per_game": 87.6875, "stones_lost_per_game": 3.125, "passes_per_game": 2.375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 32, "end_repetition": 0, "end_move_limit": 0}}, "files": "/work/pass_5/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 20.

python

exec(open("/work/analysis_helpers.py").read())games=json.load(open("/work/pass_5/games.json"))boards=np.load("/work/pass_5/boards.npy")for color in ["black","white"]:    gg=[g for g in games if g["opponent"]=="strong" and g["net_color"]==color]    print(color, "margins", sorted(g["margin"] for g in gg))for i,g in [(i,g) for i,g in enumerate(games) if g["opponent"]=="strong" and g["net_color"]=="black" and g["result"]<0][:5]:    print("\nGAME",i,"margin",g["margin"],"end",g["end"],"captures",g["stones_captured_by_net"],"lost",g["stones_lost_by_net"])    print(show(boards[i,len(g["moves"])-1]))    print("Opening",[(m["by"],m["move"]) for m in g["moves"][:24]])    print("Last",g["moves"][-6:])

exec(open("/work/analysis_helpers.py").read())games=json.load(open("/work/pass_5/games.json"))boards=np.load("/work/pass_5/boards.npy")for color in ["black","white"]:    gg=[g for g in games if g["opponent"]=="strong" and g["net_color"]==color]    print(color, "margins", sorted(g["margin"] for g in gg))for i,g in [(i,g) for i,g in enumerate(games) if g["opponent"]=="strong" and g["net_color"]=="black" and g["result"]<0][:5]:    print("\nGAME",i,"margin",g["margin"],"end",g["end"],"captures",g["stones_captured_by_net"],"lost",g["stones_lost_by_net"])    print(show(boards[i,len(g["moves"])-1]))    print("Opening",[(m["by"],m["move"]) for m in g["moves"][:24]])    print("Last",g["moves"][-6:])

OUTPUT of Python run 9 of at most 120:

black margins [-61.5, -34.5, -29.5, -29.5, -23.5, -18.5, -17.5, -17.5, -15.5, -13.5, -12.5, -11.5, -11.5, -10.5, -9.5, -9.5, -8.5, -7.5, -6.5, -6.5, -5.5, -4.5, -3.5, -1.5, -0.5, 2.5, 10.5, 30.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5]white margins [-16.5, -15.5, -13.5, -9.5, -9.5, -8.5, -8.5, -7.5, -5.5, -5.5, -5.5, -1.5, -0.5, 2.5, 5.5, 7.5, 9.5, 64.5, 85.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5]
GAME 0 margin -34.5 end two passes captures 1 lost 28OOOOOO.O..XOOOOOOOOOOOOOXXOOOOOOXXOOOOXXOXXXOOO.OOXXXOOOOOXXXXXOOOOOXXX.OOOOOX.XXOpening [('net', 60), ('opponent', 40), ('net', 59), ('opponent', 48), ('net', 58), ('opponent', 49), ('net', 50), ('opponent', 31), ('net', 70), ('opponent', 57), ('net', 41), ('opponent', 23), ('net', 51), ('opponent', 66), ('net', 32), ('opponent', 22), ('net', 24), ('opponent', 15), ('net', 33), ('opponent', 16), ('net', 42), ('opponent', 14), ('net', 25), ('opponent', 67)]Last [{'ply': 101, 'by': 'opponent', 'move': 56}, {'ply': 102, 'by': 'net', 'move': 38, 'top5': [[38, 1.0], [81, 0.0], [39, 0.0], [47, 0.0], [71, 0.0]]}, {'ply': 103, 'by': 'opponent', 'move': 12}, {'ply': 104, 'by': 'net', 'move': 39, 'top5': [[39, 0.832], [47, 0.168], [81, 0.0], [71, 0.0], [78, 0.0]]}, {'ply': 105, 'by': 'opponent', 'move': 81}, {'ply': 106, 'by': 'net', 'move': 81, 'top5': [[81, 0.9996], [71, 0.0002], [78, 0.0002], [0, 0.0], [1, 0.0]]}]
GAME 1 margin -29.5 end two passes captures 2 lost 40XOOOOOOOOXOOO.OOOOXXOOOOOOOXXXOOOOOOXXXXO.OOOXXXXOXOOOXXXOOXXOO.XXXOOOOOX.XXOOOO.Opening [('net', 56), ('opponent', 40), ('net', 60), ('opponent', 58), ('net', 39), ('opponent', 49), ('net', 48), ('opponent', 31), ('net', 41), ('opponent', 32), ('net', 50), ('opponent', 30), ('net', 59), ('opponent', 57), ('net', 66), ('opponent', 21), ('net', 42), ('opponent', 24), ('net', 47), ('opponent', 67), ('net', 38), ('opponent', 68), ('net', 29), ('opponent', 20)]Last [{'ply': 107, 'by': 'opponent', 'move': 14}, {'ply': 108, 'by': 'net', 'move': 81, 'top5': [[81, 0.8807], [36, 0.1192], [63, 0.0], [73, 0.0], [0, 0.0]]}, {'ply': 109, 'by': 'opponent', 'move': 7}, {'ply': 110, 'by': 'net', 'move': 36, 'top5': [[81, 0.8807], [36, 0.1192], [63, 0.0], [73, 0.0], [0, 0.0]]}, {'ply': 111, 'by': 'opponent', 'move': 81}, {'ply': 112, 'by': 'net', 'move': 81, 'top5': [[81, 0.9999], [63, 0.0], [73, 0.0], [0, 0.0], [1, 0.0]]}]
GAME 4 margin -7.5 end two passes captures 16 lost 20XXX.XXOO..XXXXOOOOXXXOOO.OO.XXXOOOOOX.XXO.OOOXXXOOXOOOXXXXOXOOO.X.XOOOOOX.XXXOO..Opening [('net', 24), ('opponent', 40), ('net', 39), ('opponent', 49), ('net', 59), ('opponent', 58), ('net', 60), ('opponent', 48), ('net', 50), ('opponent', 31), ('net', 41), ('opponent', 32), ('net', 57), ('opponent', 22), ('net', 47), ('opponent', 67), ('net', 56), ('opponent', 23), ('net', 66), ('opponent', 68), ('net', 30), ('opponent', 33), ('net', 42), ('opponent', 21)]Last [{'ply': 102, 'by': 'net', 'move': 50, 'top5': [[50, 0.9992], [59, 0.0004], [41, 0.0003], [81, 0.0], [3, 0.0]]}, {'ply': 103, 'by': 'opponent', 'move': 25}, {'ply': 104, 'by': 'net', 'move': 59, 'top5': [[59, 1.0], [41, 0.0], [81, 0.0], [3, 0.0], [79, 0.0]]}, {'ply': 105, 'by': 'opponent', 'move': 62}, {'ply': 106, 'by': 'net', 'move': 81, 'top5': [[81, 0.9965], [3, 0.0018], [79, 0.0013], [65, 0.0002], [37, 0.0001]]}, {'ply': 107, 'by': 'opponent', 'move': 81}]
GAME 5 margin -6.5 end two passes captures 8 lost 35XX.XOOO.XX.XXOOOOX.XXXXOOOOXXXXXOOOO.XXXOXXOOXXXXOO.OOXOOOOOOOOXXXXXOOOOX.XXOOOOOOpening [('net', 48), ('opponent', 40), ('net', 31), ('opponent', 49), ('net', 39), ('opponent', 58), ('net', 30), ('opponent', 57), ('net', 47), ('opponent', 32), ('net', 22), ('opponent', 56), ('net', 38), ('opponent', 23), ('net', 41), ('opponent', 50), ('net', 42), ('opponent', 33), ('net', 51), ('opponent', 60), ('net', 21), ('opponent', 24), ('net', 34), ('opponent', 14)]Last [{'ply': 99, 'by': 'opponent', 'move': 81}, {'ply': 100, 'by': 'net', 'move': 8, 'top5': [[81, 0.7823], [8, 0.2175], [2, 0.0001], [10, 0.0], [36, 0.0]]}, {'ply': 101, 'by': 'opponent', 'move': 81}, {'ply': 102, 'by': 'net', 'move': 17, 'top5': [[7, 0.3588], [17, 0.3588], [81, 0.2823], [2, 0.0], [10, 0.0]]}, {'ply': 103, 'by': 'opponent', 'move': 81}, {'ply': 104, 'by': 'net', 'move': 81, 'top5': [[81, 0.9998], [2, 0.0001], [10, 0.0001], [36, 0.0], [18, 0.0]]}]
GAME 7 margin -1.5 end two passes captures 25 lost 34.OOOXXXXXO.OOOXX.XOO.OOXXXXOOXOOOXX.OOXOOXX.XOXXOXXXX.OOOOOXX.XOOOOXX.X..OOOX.X.XOpening [('net', 49), ('opponent', 40), ('net', 50), ('opponent', 58), ('net', 60), ('opponent', 48), ('net', 59), ('opponent', 57), ('net', 47), ('opponent', 39), ('net', 67), ('opponent', 31), ('net', 41), ('opponent', 32), ('net', 51), ('opponent', 56), ('net', 23), ('opponent', 22), ('net', 42), ('opponent', 24), ('net', 33), ('opponent', 15), ('net', 14), ('opponent', 13)]Last [{'ply': 96, 'by': 'net', 'move': 81, 'top5': [[81, 0.9995], [16, 0.0003], [43, 0.0], [69, 0.0], [61, 0.0]]}, {'ply': 97, 'by': 'opponent', 'move': 27}, {'ply': 98, 'by': 'net', 'move': 81, 'top5': [[81, 0.9995], [16, 0.0003], [43, 0.0], [69, 0.0], [61, 0.0]]}, {'ply': 99, 'by': 'opponent', 'move': 54}, {'ply': 100, 'by': 'net', 'move': 81, 'top5': [[81, 0.9995], [16, 0.0003], [43, 0.0], [69, 0.0], [61, 0.0]]}, {'ply': 101, 'by': 'opponent', 'move': 81}]

black margins [-61.5, -34.5, -29.5, -29.5, -23.5, -18.5, -17.5, -17.5, -15.5, -13.5, -12.5, -11.5, -11.5, -10.5, -9.5, -9.5, -8.5, -7.5, -6.5, -6.5, -5.5, -4.5, -3.5, -1.5, -0.5, 2.5, 10.5, 30.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5, 73.5]white margins [-16.5, -15.5, -13.5, -9.5, -9.5, -8.5, -8.5, -7.5, -5.5, -5.5, -5.5, -1.5, -0.5, 2.5, 5.5, 7.5, 9.5, 64.5, 85.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5, 88.5]
GAME 0 margin -34.5 end two passes captures 1 lost 28OOOOOO.O..XOOOOOOOOOOOOOXXOOOOOOXXOOOOXXOXXXOOO.OOXXXOOOOOXXXXXOOOOOXXX.OOOOOX.XXOpening [('net', 60), ('opponent', 40), ('net', 59), ('opponent', 48), ('net', 58), ('opponent', 49), ('net', 50), ('opponent', 31), ('net', 70), ('opponent', 57), ('net', 41), ('opponent', 23), ('net', 51), ('opponent', 66), ('net', 32), ('opponent', 22), ('net', 24), ('opponent', 15), ('net', 33), ('opponent', 16), ('net', 42), ('opponent', 14), ('net', 25), ('opponent', 67)]Last [{'ply': 101, 'by': 'opponent', 'move': 56}, {'ply': 102, 'by': 'net', 'move': 38, 'top5': [[38, 1.0], [81, 0.0], [39, 0.0], [47, 0.0], [71, 0.0]]}, {'ply': 103, 'by': 'opponent', 'move': 12}, {'ply': 104, 'by': 'net', 'move': 39, 'top5': [[39, 0.832], [47, 0.168], [81, 0.0], [71, 0.0], [78, 0.0]]}, {'ply': 105, 'by': 'opponent', 'move': 81}, {'ply': 106, 'by': 'net', 'move': 81, 'top5': [[81, 0.9996], [71, 0.0002], [78, 0.0002], [0, 0.0], [1, 0.0]]}]
GAME 1 margin -29.5 end two passes captures 2 lost 40XOOOOOOOOXOOO.OOOOXXOOOOOOOXXXOOOOOOXXXXO.OOOXXXXOXOOOXXXOOXXOO.XXXOOOOOX.XXOOOO.Opening [('net', 56), ('opponent', 40), ('net', 60), ('opponent', 58), ('net', 39), ('opponent', 49), ('net', 48), ('opponent', 31), ('net', 41), ('opponent', 32), ('net', 50), ('opponent', 30), ('net', 59), ('opponent', 57), ('net', 66), ('opponent', 21), ('net', 42), ('opponent', 24), ('net', 47), ('opponent', 67), ('net', 38), ('opponent', 68), ('net', 29), ('opponent', 20)]Last [{'ply': 107, 'by': 'opponent', 'move': 14}, {'ply': 108, 'by': 'net', 'move': 81, 'top5': [[81, 0.8807], [36, 0.1192], [63, 0.0], [73, 0.0], [0, 0.0]]}, {'ply': 109, 'by': 'opponent', 'move': 7}, {'ply': 110, 'by': 'net', 'move': 36, 'top5': [[81, 0.8807], [36, 0.1192], [63, 0.0], [73, 0.0], [0, 0.0]]}, {'ply': 111, 'by': 'opponent', 'move': 81}, {'ply': 112, 'by': 'net', 'move': 81, 'top5': [[81, 0.9999], [63, 0.0], [73, 0.0], [0, 0.0], [1, 0.0]]}]
GAME 4 margin -7.5 end two passes captures 16 lost 20XXX.XXOO..XXXXOOOOXXXOOO.OO.XXXOOOOOX.XXO.OOOXXXOOXOOOXXXXOXOOO.X.XOOOOOX.XXXOO..Opening [('net', 24), ('opponent', 40), ('net', 39), ('opponent', 49), ('net', 59), ('opponent', 58), ('net', 60), ('opponent', 48), ('net', 50), ('opponent', 31), ('net', 41), ('opponent', 32), ('net', 57), ('opponent', 22), ('net', 47), ('opponent', 67), ('net', 56), ('opponent', 23), ('net', 66), ('opponent', 68), ('net', 30), ('opponent', 33), ('net', 42), ('opponent', 21)]Last [{'ply': 102, 'by': 'net', 'move': 50, 'top5': [[50, 0.9992], [59, 0.0004], [41, 0.0003], [81, 0.0], [3, 0.0]]}, {'ply': 103, 'by': 'opponent', 'move': 25}, {'ply': 104, 'by': 'net', 'move': 59, 'top5': [[59, 1.0], [41, 0.0], [81, 0.0], [3, 0.0], [79, 0.0]]}, {'ply': 105, 'by': 'opponent', 'move': 62}, {'ply': 106, 'by': 'net', 'move': 81, 'top5': [[81, 0.9965], [3, 0.0018], [79, 0.0013], [65, 0.0002], [37, 0.0001]]}, {'ply': 107, 'by': 'opponent', 'move': 81}]
GAME 5 margin -6.5 end two passes captures 8 lost 35XX.XOOO.XX.XXOOOOX.XXXXOOOOXXXXXOOOO.XXXOXXOOXXXXOO.OOXOOOOOOOOXXXXXOOOOX.XXOOOOOOpening [('net', 48), ('opponent', 40), ('net', 31), ('opponent', 49), ('net', 39), ('opponent', 58), ('net', 30), ('opponent', 57), ('net', 47), ('opponent', 32), ('net', 22), ('opponent', 56), ('net', 38), ('opponent', 23), ('net', 41), ('opponent', 50), ('net', 42), ('opponent', 33), ('net', 51), ('opponent', 60), ('net', 21), ('opponent', 24), ('net', 34), ('opponent', 14)]Last [{'ply': 99, 'by': 'opponent', 'move': 81}, {'ply': 100, 'by': 'net', 'move': 8, 'top5': [[81, 0.7823], [8, 0.2175], [2, 0.0001], [10, 0.0], [36, 0.0]]}, {'ply': 101, 'by': 'opponent', 'move': 81}, {'ply': 102, 'by': 'net', 'move': 17, 'top5': [[7, 0.3588], [17, 0.3588], [81, 0.2823], [2, 0.0], [10, 0.0]]}, {'ply': 103, 'by': 'opponent', 'move': 81}, {'ply': 104, 'by': 'net', 'move': 81, 'top5': [[81, 0.9998], [2, 0.0001], [10, 0.0001], [36, 0.0], [18, 0.0]]}]
GAME 7 margin -1.5 end two passes captures 25 lost 34.OOOXXXXXO.OOOXX.XOO.OOXXXXOOXOOOXX.OOXOOXX.XOXXOXXXX.OOOOOXX.XOOOOXX.X..OOOX.X.XOpening [('net', 49), ('opponent', 40), ('net', 50), ('opponent', 58), ('net', 60), ('opponent', 48), ('net', 59), ('opponent', 57), ('net', 47), ('opponent', 39), ('net', 67), ('opponent', 31), ('net', 41), ('opponent', 32), ('net', 51), ('opponent', 56), ('net', 23), ('opponent', 22), ('net', 42), ('opponent', 24), ('net', 33), ('opponent', 15), ('net', 14), ('opponent', 13)]Last [{'ply': 96, 'by': 'net', 'move': 81, 'top5': [[81, 0.9995], [16, 0.0003], [43, 0.0], [69, 0.0], [61, 0.0]]}, {'ply': 97, 'by': 'opponent', 'move': 27}, {'ply': 98, 'by': 'net', 'move': 81, 'top5': [[81, 0.9995], [16, 0.0003], [43, 0.0], [69, 0.0], [61, 0.0]]}, {'ply': 99, 'by': 'opponent', 'move': 54}, {'ply': 100, 'by': 'net', 'move': 81, 'top5': [[81, 0.9995], [16, 0.0003], [43, 0.0], [69, 0.0], [61, 0.0]]}, {'ply': 101, 'by': 'opponent', 'move': 81}]

python

import shutilshutil.copyfile("/work/design.py","/work/design_v5.py")s=open("/work/design.py").read()marker='OB[81] = P["pass_bias"]'s=s.replace(marker,r'''# Tengen is a useful universal first move on the small board.# This gate is active only on the completely empty board.B[0][486]=1for p in range(81):    W[0][3*p,486]=-1    W[0][3*p+1,486]=-1for k in range(1,6):    W[k][486,486]=1OW[486,40]=8
''' + marker)open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v3.json"))exec(s)print("Added an empty-board-only tengen opening gate.")

import shutilshutil.copyfile("/work/design.py","/work/design_v5.py")s=open("/work/design.py").read()marker='OB[81] = P["pass_bias"]'s=s.replace(marker,r'''# Tengen is a useful universal first move on the small board.# This gate is active only on the completely empty board.B[0][486]=1for p in range(81):    W[0][3*p,486]=-1    W[0][3*p+1,486]=-1for k in range(1,6):    W[k][486,486]=1OW[486,40]=8
''' + marker)open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v3.json"))exec(s)print("Added an empty-board-only tengen opening gate.")

OUTPUT of Python run 10 of at most 120:

Games pass 6 of 25 done.

{"ok": true, "pass": 6, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.59375, "wins": 57, "losses": 39, "score_as_black": 0.4791666567325592, "score_as_white": 0.7083333134651184, "mean_margin": 31.79166603088379, "mean_length": 112.65625, "stones_captured_per_game": 49.822916666666664, "stones_lost_per_game": 14.239583333333334, "passes_per_game": 2.0, "games_with_a_pass_before_move_60": 0, "end_two_passes": 90, "end_repetition": 5, "end_move_limit": 1}, "weak": {"games": 32, "score": 0.90625, "wins": 29, "losses": 3, "score_as_black": 0.8125, "score_as_white": 1.0, "mean_margin": 68.5625, "mean_length": 125.03125, "stones_captured_per_game": 87.46875, "stones_lost_per_game": 3.46875, "passes_per_game": 2.28125, "games_with_a_pass_before_move_60": 0, "end_two_passes": 31, "end_repetition": 0, "end_move_limit": 1}}, "files": "/work/pass_6/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 6, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.59375, "wins": 57, "losses": 39, "score_as_black": 0.4791666567325592, "score_as_white": 0.7083333134651184, "mean_margin": 31.79166603088379, "mean_length": 112.65625, "stones_captured_per_game": 49.822916666666664, "stones_lost_per_game": 14.239583333333334, "passes_per_game": 2.0, "games_with_a_pass_before_move_60": 0, "end_two_passes": 90, "end_repetition": 5, "end_move_limit": 1}, "weak": {"games": 32, "score": 0.90625, "wins": 29, "losses": 3, "score_as_black": 0.8125, "score_as_white": 1.0, "mean_margin": 68.5625, "mean_length": 125.03125, "stones_captured_per_game": 87.46875, "stones_lost_per_game": 3.46875, "passes_per_game": 2.28125, "games_with_a_pass_before_move_60": 0, "end_two_passes": 31, "end_repetition": 0, "end_move_limit": 1}}, "files": "/work/pass_6/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 19.

python

import shutilshutil.copyfile("/work/design.py","/work/design_v6.py")s=open("/work/design.py").read()s=s.replace('decisiveness=1.6,','decisiveness=1.6, group_pressure=1.0,')s=s.replace('        add(2,dest,unit(color,p),3)\n        B[2][dest]=-len(adj[p])','''        add(2,dest,unit(color,p),7)        B[2][dest]=-len(adj[p])-4''')s=s.replace('            add(2,dest,unit(color,q),-0.5)','''            add(2,dest,unit(color,q),-0.5)            add(2,dest,unit(2 if color==0 else 4,q),P["group_pressure"])''')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v3.json"))overrides.update(weak_friend=2.0,weak_enemy=2.5)open("/work/params_v7.json","w").write(json.dumps(overrides))exec(s)print("Pressure features now account for vulnerable connected neighbours.")print(P)

import shutilshutil.copyfile("/work/design.py","/work/design_v6.py")s=open("/work/design.py").read()s=s.replace('decisiveness=1.6,','decisiveness=1.6, group_pressure=1.0,')s=s.replace('        add(2,dest,unit(color,p),3)\n        B[2][dest]=-len(adj[p])','''        add(2,dest,unit(color,p),7)        B[2][dest]=-len(adj[p])-4''')s=s.replace('            add(2,dest,unit(color,q),-0.5)','''            add(2,dest,unit(color,q),-0.5)            add(2,dest,unit(2 if color==0 else 4,q),P["group_pressure"])''')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v3.json"))overrides.update(weak_friend=2.0,weak_enemy=2.5)open("/work/params_v7.json","w").write(json.dumps(overrides))exec(s)print("Pressure features now account for vulnerable connected neighbours.")print(P)

OUTPUT of Python run 11 of at most 120:

Pressure features now account for vulnerable connected neighbours.{'capture': 7.0, 'rescue': 6.0, 'self_atari': 9.0, 'eye': 11.0, 'crowd': 2.0, 'pass_bias': -2.0, 'friend_adj': 1.4, 'enemy_adj': 0.5, 'friend_diag': 0.6, 'enemy_diag': 0.1, 'friend_jump': 0.25, 'enemy_jump': 0.15, 'friend_knight': 0.1, 'enemy_knight': 0.05, 'seed_shared_liberty': 1.0, 'positional_scale': 2.0, 'low_liberty': 1.5, 'weak_friend': 2.0, 'weak_enemy': 2.5, 'decisiveness': 1.6, 'group_pressure': 1.0}

Pressure features now account for vulnerable connected neighbours.{'capture': 7.0, 'rescue': 6.0, 'self_atari': 9.0, 'eye': 11.0, 'crowd': 2.0, 'pass_bias': -2.0, 'friend_adj': 1.4, 'enemy_adj': 0.5, 'friend_diag': 0.6, 'enemy_diag': 0.1, 'friend_jump': 0.25, 'enemy_jump': 0.15, 'friend_knight': 0.1, 'enemy_knight': 0.05, 'seed_shared_liberty': 1.0, 'positional_scale': 2.0, 'low_liberty': 1.5, 'weak_friend': 2.0, 'weak_enemy': 2.5, 'decisiveness': 1.6, 'group_pressure': 1.0}

Games pass 7 of 25 done.

{"ok": true, "pass": 7, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.4479166567325592, "wins": 43, "losses": 53, "score_as_black": 0.3958333432674408, "score_as_white": 0.5, "mean_margin": 12.395833015441895, "mean_length": 136.65625, "stones_captured_per_game": 53.708333333333336, "stones_lost_per_game": 49.427083333333336, "passes_per_game": 1.5625, "games_with_a_pass_before_move_60": 0, "end_two_passes": 73, "end_repetition": 2, "end_move_limit": 21}, "weak": {"games": 32, "score": 0.78125, "wins": 25, "losses": 7, "score_as_black": 0.75, "score_as_white": 0.8125, "mean_margin": 46.625, "mean_length": 147.03125, "stones_captured_per_game": 88.46875, "stones_lost_per_game": 31.3125, "passes_per_game": 1.625, "games_with_a_pass_before_move_60": 0, "end_two_passes": 21, "end_repetition": 1, "end_move_limit": 10}}, "files": "/work/pass_7/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 7, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.4479166567325592, "wins": 43, "losses": 53, "score_as_black": 0.3958333432674408, "score_as_white": 0.5, "mean_margin": 12.395833015441895, "mean_length": 136.65625, "stones_captured_per_game": 53.708333333333336, "stones_lost_per_game": 49.427083333333336, "passes_per_game": 1.5625, "games_with_a_pass_before_move_60": 0, "end_two_passes": 73, "end_repetition": 2, "end_move_limit": 21}, "weak": {"games": 32, "score": 0.78125, "wins": 25, "losses": 7, "score_as_black": 0.75, "score_as_white": 0.8125, "mean_margin": 46.625, "mean_length": 147.03125, "stones_captured_per_game": 88.46875, "stones_lost_per_game": 31.3125, "passes_per_game": 1.625, "games_with_a_pass_before_move_60": 0, "end_two_passes": 21, "end_repetition": 1, "end_move_limit": 10}}, "files": "/work/pass_7/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 18.

python

import shutilshutil.copyfile("/work/design.py","/work/design_v7.py")s=open("/work/design_v6.py").read()s=s.replace('    # Capture and rescue urgency.\n    for q in adj[p]:\n        add(5,unit(0,p),unit(3,q),1)\n        add(5,unit(1,p),unit(2,q),1)','''    # Saturated tactical bonuses avoid counting the same string repeatedly.    B[5][unit(0,p)]=1    B[5][unit(1,p)]=1    for q in adj[p]:        add(5,unit(0,p),unit(3,q),-1)        add(5,unit(1,p),unit(2,q),-1)''')s=s.replace('P["capture"], P["rescue"], -P["self_atari"]',            '-P["capture"], -P["rescue"], -P["self_atari"]')s=s.replace('OB[p] = -20 + P["positional_scale"]*positional',            'OB[p] = -20 + P["capture"] + P["rescue"] + P["positional_scale"]*positional')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v3.json"))overrides.update(self_atari=16.0)open("/work/params_v8.json","w").write(json.dumps(overrides))exec(s)print("Saturated tactical urgency; strongly reject non-capturing self-atari.")

import shutilshutil.copyfile("/work/design.py","/work/design_v7.py")s=open("/work/design_v6.py").read()s=s.replace('    # Capture and rescue urgency.\n    for q in adj[p]:\n        add(5,unit(0,p),unit(3,q),1)\n        add(5,unit(1,p),unit(2,q),1)','''    # Saturated tactical bonuses avoid counting the same string repeatedly.    B[5][unit(0,p)]=1    B[5][unit(1,p)]=1    for q in adj[p]:        add(5,unit(0,p),unit(3,q),-1)        add(5,unit(1,p),unit(2,q),-1)''')s=s.replace('P["capture"], P["rescue"], -P["self_atari"]',            '-P["capture"], -P["rescue"], -P["self_atari"]')s=s.replace('OB[p] = -20 + P["positional_scale"]*positional',            'OB[p] = -20 + P["capture"] + P["rescue"] + P["positional_scale"]*positional')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v3.json"))overrides.update(self_atari=16.0)open("/work/params_v8.json","w").write(json.dumps(overrides))exec(s)print("Saturated tactical urgency; strongly reject non-capturing self-atari.")

OUTPUT of Python run 12 of at most 120:

Games pass 8 of 25 done.

{"ok": true, "pass": 8, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.5625, "wins": 54, "losses": 42, "score_as_black": 0.5208333134651184, "score_as_white": 0.6041666865348816, "mean_margin": 33.76041793823242, "mean_length": 113.29166666666667, "stones_captured_per_game": 52.510416666666664, "stones_lost_per_game": 15.604166666666666, "passes_per_game": 2.0104166666666665, "games_with_a_pass_before_move_60": 0, "end_two_passes": 87, "end_repetition": 7, "end_move_limit": 2}, "weak": {"games": 32, "score": 1.0, "wins": 32, "losses": 0, "score_as_black": 1.0, "score_as_white": 1.0, "mean_margin": 81.0, "mean_length": 130.9375, "stones_captured_per_game": 99.71875, "stones_lost_per_game": 1.53125, "passes_per_game": 2.71875, "games_with_a_pass_before_move_60": 0, "end_two_passes": 32, "end_repetition": 0, "end_move_limit": 0}}, "files": "/work/pass_8/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 8, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.5625, "wins": 54, "losses": 42, "score_as_black": 0.5208333134651184, "score_as_white": 0.6041666865348816, "mean_margin": 33.76041793823242, "mean_length": 113.29166666666667, "stones_captured_per_game": 52.510416666666664, "stones_lost_per_game": 15.604166666666666, "passes_per_game": 2.0104166666666665, "games_with_a_pass_before_move_60": 0, "end_two_passes": 87, "end_repetition": 7, "end_move_limit": 2}, "weak": {"games": 32, "score": 1.0, "wins": 32, "losses": 0, "score_as_black": 1.0, "score_as_white": 1.0, "mean_margin": 81.0, "mean_length": 130.9375, "stones_captured_per_game": 99.71875, "stones_lost_per_game": 1.53125, "passes_per_game": 2.71875, "games_with_a_pass_before_move_60": 0, "end_two_passes": 32, "end_repetition": 0, "end_move_limit": 0}}, "files": "/work/pass_8/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 17.

python

import shutilshutil.copyfile("/work/design.py","/work/design_v8.py")s=open("/work/design.py").read()a=s.index('    # Prefer moves with room')b=s.index('# Fifth layer',a)s=s[:a]+r'''    # Binary access to a friendly string with spare liberties.    # Count the string once even if it touches a move on several sides.    dest=unit(5,p)    B[3][dest]=1    for q in adj[p]:        add(3,dest,unit(0,q),-1)        add(3,dest,unit(2,q),1)
''' + s[b:]a=s.index('    # Insufficient escape')b=s.index('    dest=unit(3,p)',a)s=s[:a]+r'''    # Conservative self-atari prevention with saturated friendly access.    dest=unit(2,p)    B[5][dest]=0.5-len(adj[p])    add(5,dest,unit(5,p),1.5)    for q in adj[p]:        add(5,dest,unit(0,q),1)        add(5,dest,unit(1,q),1)        add(5,dest,unit(3,q),-2)
''' + s[b:]s=s.replace('    for q in diag[p]:\n        add(5,dest,unit(0,q),0.5)','''    for q in adj[p]:        add(5,dest,unit(1,q),-0.75)    for q in diag[p]:        add(5,dest,unit(0,q),0.5)''')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v8.json"))overrides.update(low_liberty=0.5)open("/work/params_v9.json","w").write(json.dumps(overrides))exec(s)print("Conservative string access; permit necessary connections in contested areas.")

import shutilshutil.copyfile("/work/design.py","/work/design_v8.py")s=open("/work/design.py").read()a=s.index('    # Prefer moves with room')b=s.index('# Fifth layer',a)s=s[:a]+r'''    # Binary access to a friendly string with spare liberties.    # Count the string once even if it touches a move on several sides.    dest=unit(5,p)    B[3][dest]=1    for q in adj[p]:        add(3,dest,unit(0,q),-1)        add(3,dest,unit(2,q),1)
''' + s[b:]a=s.index('    # Insufficient escape')b=s.index('    dest=unit(3,p)',a)s=s[:a]+r'''    # Conservative self-atari prevention with saturated friendly access.    dest=unit(2,p)    B[5][dest]=0.5-len(adj[p])    add(5,dest,unit(5,p),1.5)    for q in adj[p]:        add(5,dest,unit(0,q),1)        add(5,dest,unit(1,q),1)        add(5,dest,unit(3,q),-2)
''' + s[b:]s=s.replace('    for q in diag[p]:\n        add(5,dest,unit(0,q),0.5)','''    for q in adj[p]:        add(5,dest,unit(1,q),-0.75)    for q in diag[p]:        add(5,dest,unit(0,q),0.5)''')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v8.json"))overrides.update(low_liberty=0.5)open("/work/params_v9.json","w").write(json.dumps(overrides))exec(s)print("Conservative string access; permit necessary connections in contested areas.")

OUTPUT of Python run 13 of at most 120:

Games pass 9 of 25 done.

{"ok": true, "pass": 9, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.65625, "wins": 63, "losses": 33, "score_as_black": 0.5625, "score_as_white": 0.75, "mean_margin": 41.32291793823242, "mean_length": 116.625, "stones_captured_per_game": 60.9375, "stones_lost_per_game": 14.03125, "passes_per_game": 1.8020833333333333, "games_with_a_pass_before_move_60": 0, "end_two_passes": 84, "end_repetition": 8, "end_move_limit": 4}, "weak": {"games": 32, "score": 1.0, "wins": 32, "losses": 0, "score_as_black": 1.0, "score_as_white": 1.0, "mean_margin": 72.4375, "mean_length": 121.03125, "stones_captured_per_game": 85.1875, "stones_lost_per_game": 2.21875, "passes_per_game": 2.53125, "games_with_a_pass_before_move_60": 0, "end_two_passes": 31, "end_repetition": 1, "end_move_limit": 0}}, "files": "/work/pass_9/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 9, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.65625, "wins": 63, "losses": 33, "score_as_black": 0.5625, "score_as_white": 0.75, "mean_margin": 41.32291793823242, "mean_length": 116.625, "stones_captured_per_game": 60.9375, "stones_lost_per_game": 14.03125, "passes_per_game": 1.8020833333333333, "games_with_a_pass_before_move_60": 0, "end_two_passes": 84, "end_repetition": 8, "end_move_limit": 4}, "weak": {"games": 32, "score": 1.0, "wins": 32, "losses": 0, "score_as_black": 1.0, "score_as_white": 1.0, "mean_margin": 72.4375, "mean_length": 121.03125, "stones_captured_per_game": 85.1875, "stones_lost_per_game": 2.21875, "passes_per_game": 2.53125, "games_with_a_pass_before_move_60": 0, "end_two_passes": 31, "end_repetition": 1, "end_move_limit": 0}}, "files": "/work/pass_9/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 16.

python

import shutilshutil.copyfile("/work/design.py","/work/design_v9.py")s=open("/work/design.py").read()s=s.replace('    add(5,dest,unit(5,p),-P["low_liberty"])',r'''    # Extend from healthy strings rather than from stones in trouble.    for offsets, strength in [(DI,0.3),(JU,0.5),(KN,0.2)]:        for q in points(p,offsets):            add(5,dest,unit(0,q),strength)            add(5,dest,unit(2,q),-strength)    add(5,dest,unit(5,p),-P["low_liberty"])''')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added healthy-string influence for diagonal, one-space, and knight extensions.")

import shutilshutil.copyfile("/work/design.py","/work/design_v9.py")s=open("/work/design.py").read()s=s.replace('    add(5,dest,unit(5,p),-P["low_liberty"])',r'''    # Extend from healthy strings rather than from stones in trouble.    for offsets, strength in [(DI,0.3),(JU,0.5),(KN,0.2)]:        for q in points(p,offsets):            add(5,dest,unit(0,q),strength)            add(5,dest,unit(2,q),-strength)    add(5,dest,unit(5,p),-P["low_liberty"])''')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added healthy-string influence for diagonal, one-space, and knight extensions.")

OUTPUT of Python run 14 of at most 120:

Games pass 10 of 25 done.

{"ok": true, "pass": 10, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.5416666865348816, "wins": 52, "losses": 44, "score_as_black": 0.3958333432674408, "score_as_white": 0.6875, "mean_margin": 33.32291793823242, "mean_length": 110.3125, "stones_captured_per_game": 50.604166666666664, "stones_lost_per_game": 10.791666666666666, "passes_per_game": 2.03125, "games_with_a_pass_before_move_60": 0, "end_two_passes": 88, "end_repetition": 8, "end_move_limit": 0}, "weak": {"games": 32, "score": 0.875, "wins": 28, "losses": 4, "score_as_black": 0.8125, "score_as_white": 0.9375, "mean_margin": 62.46875, "mean_length": 119.625, "stones_captured_per_game": 79.0625, "stones_lost_per_game": 4.1875, "passes_per_game": 3.0625, "games_with_a_pass_before_move_60": 0, "end_two_passes": 30, "end_repetition": 2, "end_move_limit": 0}}, "files": "/work/pass_10/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 10, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.5416666865348816, "wins": 52, "losses": 44, "score_as_black": 0.3958333432674408, "score_as_white": 0.6875, "mean_margin": 33.32291793823242, "mean_length": 110.3125, "stones_captured_per_game": 50.604166666666664, "stones_lost_per_game": 10.791666666666666, "passes_per_game": 2.03125, "games_with_a_pass_before_move_60": 0, "end_two_passes": 88, "end_repetition": 8, "end_move_limit": 0}, "weak": {"games": 32, "score": 0.875, "wins": 28, "losses": 4, "score_as_black": 0.8125, "score_as_white": 0.9375, "mean_margin": 62.46875, "mean_length": 119.625, "stones_captured_per_game": 79.0625, "stones_lost_per_game": 4.1875, "passes_per_game": 3.0625, "games_with_a_pass_before_move_60": 0, "end_two_passes": 30, "end_repetition": 2, "end_move_limit": 0}}, "files": "/work/pass_10/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 15.

python

import json,numpy as npexec(open("/work/analysis_helpers.py").read())gs=json.load(open("/work/pass_9/games.json")); bs=np.load("/work/pass_9/boards.npy")for i,g in enumerate(gs):    if g["end"]!="repetition": continue    print("\nGAME",i,g["opponent"],g["net_color"],"result",g["result"],"margin",g["margin"])    print(show(bs[i,len(g["moves"])-2]))    print(g["moves"][-7:])

import json,numpy as npexec(open("/work/analysis_helpers.py").read())gs=json.load(open("/work/pass_9/games.json")); bs=np.load("/work/pass_9/boards.npy")for i,g in enumerate(gs):    if g["end"]!="repetition": continue    print("\nGAME",i,g["opponent"],g["net_color"],"result",g["result"],"margin",g["margin"])    print(show(bs[i,len(g["moves"])-2]))    print(g["moves"][-7:])

OUTPUT of Python run 15 of at most 120:


GAME 12 strong black result -1 margin -24.5...OOOO.OOOOOOOO..XXOO..OOO.XXOOOOOOXXXXXXOOO.XXXOOO.OXXXXO..OOXXXXXOOOO.X.XXXOO.[{'ply': 90, 'by': 'net', 'move': 22, 'top5': [[22, 0.6551], [14, 0.1973], [1, 0.1433], [17, 0.0015], [7, 0.0014]]}, {'ply': 91, 'by': 'opponent', 'move': 14}, {'ply': 92, 'by': 'net', 'move': 17, 'top5': [[17, 1.0], [1, 0.0], [7, 0.0], [81, 0.0], [59, 0.0]]}, {'ply': 93, 'by': 'opponent', 'move': 81}, {'ply': 94, 'by': 'net', 'move': 7, 'top5': [[7, 1.0], [1, 0.0], [81, 0.0], [59, 0.0], [22, 0.0]]}, {'ply': 95, 'by': 'opponent', 'move': 8}, {'ply': 96, 'by': 'net', 'move': 16, 'top5': [[16, 0.9832], [1, 0.0166], [81, 0.0001], [59, 0.0], [22, 0.0]]}]
GAME 27 strong black result -1 margin 6.5OX.X.XXO..OXXXXO.O.OOOOOOOOOXXOXXXXOOOXOX.XXXOOXX.XXX.OOX.X.X.XOOOXXXXXX.OO......[{'ply': 74, 'by': 'net', 'move': 71, 'top5': [[0, 0.3025], [76, 0.1872], [78, 0.1728], [71, 0.1255], [75, 0.0987]]}, {'ply': 75, 'by': 'opponent', 'move': 0}, {'ply': 76, 'by': 'net', 'move': 9, 'top5': [[9, 0.9427], [18, 0.0573], [2, 0.0], [78, 0.0], [76, 0.0]]}, {'ply': 77, 'by': 'opponent', 'move': 74}, {'ply': 78, 'by': 'net', 'move': 18, 'top5': [[18, 1.0], [75, 0.0], [76, 0.0], [78, 0.0], [80, 0.0]]}, {'ply': 79, 'by': 'opponent', 'move': 0}, {'ply': 80, 'by': 'net', 'move': 9, 'top5': [[9, 0.9427], [18, 0.0573], [75, 0.0], [2, 0.0], [76, 0.0]]}]
GAME 29 strong black result -1 margin 8.5XXXOOOOO..XOOOXOOOX.XOOXXXO.XXOXXX.XXXXOX.XX.X.XOOXXXXXXXXOXXXOXXOOOOXO.XOOOOOOO.[{'ply': 84, 'by': 'net', 'move': 63, 'top5': [[63, 0.9997], [81, 0.0003], [80, 0.0], [62, 0.0], [41, 0.0]]}, {'ply': 85, 'by': 'opponent', 'move': 76}, {'ply': 86, 'by': 'net', 'move': 81, 'top5': [[81, 0.9693], [80, 0.0192], [62, 0.011], [41, 0.0004], [34, 0.0]]}, {'ply': 87, 'by': 'opponent', 'move': 77}, {'ply': 88, 'by': 'net', 'move': 80, 'top5': [[80, 1.0], [81, 0.0], [62, 0.0], [41, 0.0], [34, 0.0]]}, {'ply': 89, 'by': 'opponent', 'move': 62}, {'ply': 90, 'by': 'net', 'move': 71, 'top5': [[71, 1.0], [81, 0.0], [41, 0.0], [34, 0.0], [19, 0.0]]}]
GAME 51 strong white result -1 margin -3.5.X.X.XXX.XX..OXXX.XXX.XXXOXXXXXXOOOOXXO.XOOO..X.XOOOOOX.XXXOOOOXXXXOOOOO.XXXXO.O.[{'ply': 91, 'by': 'net', 'move': 13, 'top5': [[39, 0.4435], [13, 0.2159], [11, 0.184], [26, 0.1336], [12, 0.0212]]}, {'ply': 92, 'by': 'opponent', 'move': 26}, {'ply': 93, 'by': 'net', 'move': 17, 'top5': [[17, 1.0], [12, 0.0], [39, 0.0], [11, 0.0], [47, 0.0]]}, {'ply': 94, 'by': 'opponent', 'move': 73}, {'ply': 95, 'by': 'net', 'move': 8, 'top5': [[4, 0.7231], [8, 0.2769], [12, 0.0], [39, 0.0], [11, 0.0]]}, {'ply': 96, 'by': 'opponent', 'move': 26}, {'ply': 97, 'by': 'net', 'move': 17, 'top5': [[17, 1.0], [12, 0.0], [39, 0.0], [11, 0.0], [47, 0.0]]}]
GAME 54 strong white result -1 margin 2.5O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOOXXXXXXXX.XXXXXX.[{'ply': 93, 'by': 'net', 'move': 81, 'top5': [[81, 0.996], [59, 0.004], [60, 0.0], [11, 0.0], [1, 0.0]]}, {'ply': 94, 'by': 'opponent', 'move': 72}, {'ply': 95, 'by': 'net', 'move': 64, 'top5': [[64, 0.9999], [81, 0.0001], [59, 0.0], [60, 0.0], [63, 0.0]]}, {'ply': 96, 'by': 'opponent', 'move': 81}, {'ply': 97, 'by': 'net', 'move': 63, 'top5': [[63, 0.5], [73, 0.5], [81, 0.0], [59, 0.0], [60, 0.0]]}, {'ply': 98, 'by': 'opponent', 'move': 81}, {'ply': 99, 'by': 'net', 'move': 73, 'top5': [[73, 1.0], [81, 0.0], [59, 0.0], [60, 0.0], [11, 0.0]]}]
GAME 57 strong white result -1 margin 33.5.O.OOOO.OOOOOO.OOOOO.OOOOOX.OOO.OOXXO.O.OOOXXOOOOOOOXXXOXXOOOXX.XXXXXXXX.XXXX.XX.[{'ply': 89, 'by': 'net', 'move': 72, 'top5': [[72, 0.9991], [81, 0.0009], [54, 0.0001], [20, 0.0], [31, 0.0]]}, {'ply': 90, 'by': 'opponent', 'move': 54}, {'ply': 91, 'by': 'net', 'move': 63, 'top5': [[63, 1.0], [72, 0.0], [81, 0.0], [73, 0.0], [20, 0.0]]}, {'ply': 92, 'by': 'opponent', 'move': 73}, {'ply': 93, 'by': 'net', 'move': 72, 'top5': [[72, 1.0], [81, 0.0], [54, 0.0], [20, 0.0], [31, 0.0]]}, {'ply': 94, 'by': 'opponent', 'move': 54}, {'ply': 95, 'by': 'net', 'move': 63, 'top5': [[63, 1.0], [81, 0.0], [20, 0.0], [31, 0.0], [39, 0.0]]}]
GAME 58 strong white result -1 margin 13.5.OX..XO..O.OXXXXO..OXXXOOO.O.OXXOO.X.OOXXOOOXO.OXXOOOO.OOXXOOO.OXXXXXXXX..X......[{'ply': 67, 'by': 'net', 'move': 53, 'top5': [[53, 0.9645], [4, 0.0128], [6, 0.0073], [26, 0.0068], [62, 0.0036]]}, {'ply': 68, 'by': 'opponent', 'move': 71}, {'ply': 69, 'by': 'net', 'move': 6, 'top5': [[4, 0.3974], [6, 0.227], [26, 0.2096], [5, 0.0333], [17, 0.0284]]}, {'ply': 70, 'by': 'opponent', 'move': 5}, {'ply': 71, 'by': 'net', 'move': 4, 'top5': [[4, 0.9953], [7, 0.0047], [26, 0.0], [17, 0.0], [72, 0.0]]}, {'ply': 72, 'by': 'opponent', 'move': 2}, {'ply': 73, 'by': 'net', 'move': 3, 'top5': [[3, 0.9912], [7, 0.0088], [10, 0.0], [26, 0.0], [17, 0.0]]}]
GAME 80 strong white result -1 margin -2.5.XXO.XX..XXXOXXXXOXXXOXXXO.XXXXXXOOO.X.XXOOO.XXOXOOOOOXXOXXOOO.XXXXXOOOO.XXXOO.OO[{'ply': 121, 'by': 'net', 'move': 21, 'top5': [[21, 0.9549], [7, 0.0422], [4, 0.0016], [81, 0.0011], [2, 0.0003]]}, {'ply': 122, 'by': 'opponent', 'move': 2}, {'ply': 123, 'by': 'net', 'move': 7, 'top5': [[7, 0.9739], [81, 0.0246], [26, 0.0015], [62, 0.0], [44, 0.0]]}, {'ply': 124, 'by': 'opponent', 'move': 26}, {'ply': 125, 'by': 'net', 'move': 17, 'top5': [[17, 1.0], [8, 0.0], [81, 0.0], [7, 0.0], [44, 0.0]]}, {'ply': 126, 'by': 'opponent', 'move': 81}, {'ply': 127, 'by': 'net', 'move': 8, 'top5': [[8, 0.9917], [81, 0.0082], [26, 0.0001], [7, 0.0], [62, 0.0]]}]
GAME 105 weak black result 1 margin 2.5.XXX.XXX.XX.XXXOOOX.XXOOOO..XXOOOOOX.XXOXXOXXXOXXXOXXXOOOXXOX.XOOOOOOXX..O..OOX.X[{'ply': 69, 'by': 'opponent', 'move': 54}, {'ply': 70, 'by': 'net', 'move': 35, 'top5': [[35, 0.9703], [7, 0.0287], [26, 0.0008], [17, 0.0001], [8, 0.0]]}, {'ply': 71, 'by': 'opponent', 'move': 17}, {'ply': 72, 'by': 'net', 'move': 7, 'top5': [[7, 0.9997], [81, 0.0002], [27, 0.0001], [75, 0.0], [74, 0.0]]}, {'ply': 73, 'by': 'opponent', 'move': 27}, {'ply': 74, 'by': 'net', 'move': 45, 'top5': [[45, 1.0], [81, 0.0], [19, 0.0], [75, 0.0], [74, 0.0]]}, {'ply': 75, 'by': 'opponent', 'move': 36}]


GAME 12 strong black result -1 margin -24.5...OOOO.OOOOOOOO..XXOO..OOO.XXOOOOOOXXXXXXOOO.XXXOOO.OXXXXO..OOXXXXXOOOO.X.XXXOO.[{'ply': 90, 'by': 'net', 'move': 22, 'top5': [[22, 0.6551], [14, 0.1973], [1, 0.1433], [17, 0.0015], [7, 0.0014]]}, {'ply': 91, 'by': 'opponent', 'move': 14}, {'ply': 92, 'by': 'net', 'move': 17, 'top5': [[17, 1.0], [1, 0.0], [7, 0.0], [81, 0.0], [59, 0.0]]}, {'ply': 93, 'by': 'opponent', 'move': 81}, {'ply': 94, 'by': 'net', 'move': 7, 'top5': [[7, 1.0], [1, 0.0], [81, 0.0], [59, 0.0], [22, 0.0]]}, {'ply': 95, 'by': 'opponent', 'move': 8}, {'ply': 96, 'by': 'net', 'move': 16, 'top5': [[16, 0.9832], [1, 0.0166], [81, 0.0001], [59, 0.0], [22, 0.0]]}]
GAME 27 strong black result -1 margin 6.5OX.X.XXO..OXXXXO.O.OOOOOOOOOXXOXXXXOOOXOX.XXXOOXX.XXX.OOX.X.X.XOOOXXXXXX.OO......[{'ply': 74, 'by': 'net', 'move': 71, 'top5': [[0, 0.3025], [76, 0.1872], [78, 0.1728], [71, 0.1255], [75, 0.0987]]}, {'ply': 75, 'by': 'opponent', 'move': 0}, {'ply': 76, 'by': 'net', 'move': 9, 'top5': [[9, 0.9427], [18, 0.0573], [2, 0.0], [78, 0.0], [76, 0.0]]}, {'ply': 77, 'by': 'opponent', 'move': 74}, {'ply': 78, 'by': 'net', 'move': 18, 'top5': [[18, 1.0], [75, 0.0], [76, 0.0], [78, 0.0], [80, 0.0]]}, {'ply': 79, 'by': 'opponent', 'move': 0}, {'ply': 80, 'by': 'net', 'move': 9, 'top5': [[9, 0.9427], [18, 0.0573], [75, 0.0], [2, 0.0], [76, 0.0]]}]
GAME 29 strong black result -1 margin 8.5XXXOOOOO..XOOOXOOOX.XOOXXXO.XXOXXX.XXXXOX.XX.X.XOOXXXXXXXXOXXXOXXOOOOXO.XOOOOOOO.[{'ply': 84, 'by': 'net', 'move': 63, 'top5': [[63, 0.9997], [81, 0.0003], [80, 0.0], [62, 0.0], [41, 0.0]]}, {'ply': 85, 'by': 'opponent', 'move': 76}, {'ply': 86, 'by': 'net', 'move': 81, 'top5': [[81, 0.9693], [80, 0.0192], [62, 0.011], [41, 0.0004], [34, 0.0]]}, {'ply': 87, 'by': 'opponent', 'move': 77}, {'ply': 88, 'by': 'net', 'move': 80, 'top5': [[80, 1.0], [81, 0.0], [62, 0.0], [41, 0.0], [34, 0.0]]}, {'ply': 89, 'by': 'opponent', 'move': 62}, {'ply': 90, 'by': 'net', 'move': 71, 'top5': [[71, 1.0], [81, 0.0], [41, 0.0], [34, 0.0], [19, 0.0]]}]
GAME 51 strong white result -1 margin -3.5.X.X.XXX.XX..OXXX.XXX.XXXOXXXXXXOOOOXXO.XOOO..X.XOOOOOX.XXXOOOOXXXXOOOOO.XXXXO.O.[{'ply': 91, 'by': 'net', 'move': 13, 'top5': [[39, 0.4435], [13, 0.2159], [11, 0.184], [26, 0.1336], [12, 0.0212]]}, {'ply': 92, 'by': 'opponent', 'move': 26}, {'ply': 93, 'by': 'net', 'move': 17, 'top5': [[17, 1.0], [12, 0.0], [39, 0.0], [11, 0.0], [47, 0.0]]}, {'ply': 94, 'by': 'opponent', 'move': 73}, {'ply': 95, 'by': 'net', 'move': 8, 'top5': [[4, 0.7231], [8, 0.2769], [12, 0.0], [39, 0.0], [11, 0.0]]}, {'ply': 96, 'by': 'opponent', 'move': 26}, {'ply': 97, 'by': 'net', 'move': 17, 'top5': [[17, 1.0], [12, 0.0], [39, 0.0], [11, 0.0], [47, 0.0]]}]
GAME 54 strong white result -1 margin 2.5O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOOXXXXXXXX.XXXXXX.[{'ply': 93, 'by': 'net', 'move': 81, 'top5': [[81, 0.996], [59, 0.004], [60, 0.0], [11, 0.0], [1, 0.0]]}, {'ply': 94, 'by': 'opponent', 'move': 72}, {'ply': 95, 'by': 'net', 'move': 64, 'top5': [[64, 0.9999], [81, 0.0001], [59, 0.0], [60, 0.0], [63, 0.0]]}, {'ply': 96, 'by': 'opponent', 'move': 81}, {'ply': 97, 'by': 'net', 'move': 63, 'top5': [[63, 0.5], [73, 0.5], [81, 0.0], [59, 0.0], [60, 0.0]]}, {'ply': 98, 'by': 'opponent', 'move': 81}, {'ply': 99, 'by': 'net', 'move': 73, 'top5': [[73, 1.0], [81, 0.0], [59, 0.0], [60, 0.0], [11, 0.0]]}]
GAME 57 strong white result -1 margin 33.5.O.OOOO.OOOOOO.OOOOO.OOOOOX.OOO.OOXXO.O.OOOXXOOOOOOOXXXOXXOOOXX.XXXXXXXX.XXXX.XX.[{'ply': 89, 'by': 'net', 'move': 72, 'top5': [[72, 0.9991], [81, 0.0009], [54, 0.0001], [20, 0.0], [31, 0.0]]}, {'ply': 90, 'by': 'opponent', 'move': 54}, {'ply': 91, 'by': 'net', 'move': 63, 'top5': [[63, 1.0], [72, 0.0], [81, 0.0], [73, 0.0], [20, 0.0]]}, {'ply': 92, 'by': 'opponent', 'move': 73}, {'ply': 93, 'by': 'net', 'move': 72, 'top5': [[72, 1.0], [81, 0.0], [54, 0.0], [20, 0.0], [31, 0.0]]}, {'ply': 94, 'by': 'opponent', 'move': 54}, {'ply': 95, 'by': 'net', 'move': 63, 'top5': [[63, 1.0], [81, 0.0], [20, 0.0], [31, 0.0], [39, 0.0]]}]
GAME 58 strong white result -1 margin 13.5.OX..XO..O.OXXXXO..OXXXOOO.O.OXXOO.X.OOXXOOOXO.OXXOOOO.OOXXOOO.OXXXXXXXX..X......[{'ply': 67, 'by': 'net', 'move': 53, 'top5': [[53, 0.9645], [4, 0.0128], [6, 0.0073], [26, 0.0068], [62, 0.0036]]}, {'ply': 68, 'by': 'opponent', 'move': 71}, {'ply': 69, 'by': 'net', 'move': 6, 'top5': [[4, 0.3974], [6, 0.227], [26, 0.2096], [5, 0.0333], [17, 0.0284]]}, {'ply': 70, 'by': 'opponent', 'move': 5}, {'ply': 71, 'by': 'net', 'move': 4, 'top5': [[4, 0.9953], [7, 0.0047], [26, 0.0], [17, 0.0], [72, 0.0]]}, {'ply': 72, 'by': 'opponent', 'move': 2}, {'ply': 73, 'by': 'net', 'move': 3, 'top5': [[3, 0.9912], [7, 0.0088], [10, 0.0], [26, 0.0], [17, 0.0]]}]
GAME 80 strong white result -1 margin -2.5.XXO.XX..XXXOXXXXOXXXOXXXO.XXXXXXOOO.X.XXOOO.XXOXOOOOOXXOXXOOO.XXXXXOOOO.XXXOO.OO[{'ply': 121, 'by': 'net', 'move': 21, 'top5': [[21, 0.9549], [7, 0.0422], [4, 0.0016], [81, 0.0011], [2, 0.0003]]}, {'ply': 122, 'by': 'opponent', 'move': 2}, {'ply': 123, 'by': 'net', 'move': 7, 'top5': [[7, 0.9739], [81, 0.0246], [26, 0.0015], [62, 0.0], [44, 0.0]]}, {'ply': 124, 'by': 'opponent', 'move': 26}, {'ply': 125, 'by': 'net', 'move': 17, 'top5': [[17, 1.0], [8, 0.0], [81, 0.0], [7, 0.0], [44, 0.0]]}, {'ply': 126, 'by': 'opponent', 'move': 81}, {'ply': 127, 'by': 'net', 'move': 8, 'top5': [[8, 0.9917], [81, 0.0082], [26, 0.0001], [7, 0.0], [62, 0.0]]}]
GAME 105 weak black result 1 margin 2.5.XXX.XXX.XX.XXXOOOX.XXOOOO..XXOOOOOX.XXOXXOXXXOXXXOXXXOOOXXOX.XOOOOOOXX..O..OOX.X[{'ply': 69, 'by': 'opponent', 'move': 54}, {'ply': 70, 'by': 'net', 'move': 35, 'top5': [[35, 0.9703], [7, 0.0287], [26, 0.0008], [17, 0.0001], [8, 0.0]]}, {'ply': 71, 'by': 'opponent', 'move': 17}, {'ply': 72, 'by': 'net', 'move': 7, 'top5': [[7, 0.9997], [81, 0.0002], [27, 0.0001], [75, 0.0], [74, 0.0]]}, {'ply': 73, 'by': 'opponent', 'move': 27}, {'ply': 74, 'by': 'net', 'move': 45, 'top5': [[45, 1.0], [81, 0.0], [19, 0.0], [75, 0.0], [74, 0.0]]}, {'ply': 75, 'by': 'opponent', 'move': 36}]

python

g=gs[54]; i=54for t in range(89,len(g["moves"])):    print("\n",g["moves"][t])    print(show(bs[i,t]))

g=gs[54]; i=54for t in range(89,len(g["moves"])):    print("\n",g["moves"][t])    print(show(bs[i,t]))

OUTPUT of Python run 16 of at most 120:


[stderr]Traceback (most recent call last):  File "<string>", line 32, in <module>NameError: name 'gs' is not defined. Did you mean: 'os'?


[stderr]Traceback (most recent call last):  File "<string>", line 32, in <module>NameError: name 'gs' is not defined. Did you mean: 'os'?

python

import json,numpy as npexec(open("/work/analysis_helpers.py").read())gs=json.load(open("/work/pass_9/games.json")); bs=np.load("/work/pass_9/boards.npy")g=gs[54]; i=54for t in range(89,len(g["moves"])):    print("\n",g["moves"][t])    print(show(bs[i,t]))

import json,numpy as npexec(open("/work/analysis_helpers.py").read())gs=json.load(open("/work/pass_9/games.json")); bs=np.load("/work/pass_9/boards.npy")g=gs[54]; i=54for t in range(89,len(g["moves"])):    print("\n",g["moves"][t])    print(show(bs[i,t]))

OUTPUT of Python run 17 of at most 120:


 {'ply': 89, 'by': 'net', 'move': 63, 'top5': [[63, 1.0], [65, 0.0], [81, 0.0], [73, 0.0], [59, 0.0]]}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOO.XXXXXXX.XXXXXX.
 {'ply': 90, 'by': 'opponent', 'move': 81}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOO.XXXXXXX.XXXXXX.
 {'ply': 91, 'by': 'net', 'move': 73, 'top5': [[73, 1.0], [65, 0.0], [81, 0.0], [59, 0.0], [60, 0.0]]}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOO.XXXXXX.OXXXXXX.
 {'ply': 92, 'by': 'opponent', 'move': 65}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOOXXXXXXX.OXXXXXX.
 {'ply': 93, 'by': 'net', 'move': 81, 'top5': [[81, 0.996], [59, 0.004], [60, 0.0], [11, 0.0], [1, 0.0]]}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOOXXXXXXX.OXXXXXX.
 {'ply': 94, 'by': 'opponent', 'move': 72}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XX..XXXXXXXX.XXXXXX.
 {'ply': 95, 'by': 'net', 'move': 64, 'top5': [[64, 0.9999], [81, 0.0001], [59, 0.0], [60, 0.0], [63, 0.0]]}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XX.OXXXXXXXX.XXXXXX.
 {'ply': 96, 'by': 'opponent', 'move': 81}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XX.OXXXXXXXX.XXXXXX.
 {'ply': 97, 'by': 'net', 'move': 63, 'top5': [[63, 0.5], [73, 0.5], [81, 0.0], [59, 0.0], [60, 0.0]]}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOOXXXXXXXX.XXXXXX.
 {'ply': 98, 'by': 'opponent', 'move': 81}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOOXXXXXXXX.XXXXXX.
 {'ply': 99, 'by': 'net', 'move': 73, 'top5': [[73, 1.0], [81, 0.0], [59, 0.0], [60, 0.0], [11, 0.0]]}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOOXXXXXXX.OXXXXXX.


 {'ply': 89, 'by': 'net', 'move': 63, 'top5': [[63, 1.0], [65, 0.0], [81, 0.0], [73, 0.0], [59, 0.0]]}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOO.XXXXXXX.XXXXXX.
 {'ply': 90, 'by': 'opponent', 'move': 81}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOO.XXXXXXX.XXXXXX.
 {'ply': 91, 'by': 'net', 'move': 73, 'top5': [[73, 1.0], [65, 0.0], [81, 0.0], [59, 0.0], [60, 0.0]]}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOO.XXXXXX.OXXXXXX.
 {'ply': 92, 'by': 'opponent', 'move': 65}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOOXXXXXXX.OXXXXXX.
 {'ply': 93, 'by': 'net', 'move': 81, 'top5': [[81, 0.996], [59, 0.004], [60, 0.0], [11, 0.0], [1, 0.0]]}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOOXXXXXXX.OXXXXXX.
 {'ply': 94, 'by': 'opponent', 'move': 72}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XX..XXXXXXXX.XXXXXX.
 {'ply': 95, 'by': 'net', 'move': 64, 'top5': [[64, 0.9999], [81, 0.0001], [59, 0.0], [60, 0.0], [63, 0.0]]}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XX.OXXXXXXXX.XXXXXX.
 {'ply': 96, 'by': 'opponent', 'move': 81}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XX.OXXXXXXXX.XXXXXX.
 {'ply': 97, 'by': 'net', 'move': 63, 'top5': [[63, 0.5], [73, 0.5], [81, 0.0], [59, 0.0], [60, 0.0]]}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOOXXXXXXXX.XXXXXX.
 {'ply': 98, 'by': 'opponent', 'move': 81}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOOXXXXXXXX.XXXXXX.
 {'ply': 99, 'by': 'net', 'move': 73, 'top5': [[73, 1.0], [81, 0.0], [59, 0.0], [60, 0.0], [11, 0.0]]}O.OO.O.OOOO.OO.OOXOOOOOOOXXOXOOOOXXXOXXOOOXXXXXXXOXXXXXXXXX..XXOOXXXXXXX.OXXXXXX.

python

import json,shutilshutil.copyfile("/work/design.py","/work/design_v10.py")s=open("/work/design_v9.py").read()a=s.index('    # Local liberty pressure')b=s.index('# Fourth layer',a)s=s[:a]+r'''    # Long-range safety messages travel only through proven same-colour    # bridges: a stone and ALL its neighbours must have that colour.    # Every Manhattan-distance-two endpoint is then connected to it.    ring2=points(p,DI+JU)    for color,ach,ch in [(0,2,4),(1,4,5)]:        dest=unit(ch,p)        B[2][dest]=-len(adj[p])-1        add(2,dest,unit(color,p),1)        for q in adj[p]:            add(2,dest,unit(color,q),1)        for q in ring2:            add(2,dest,unit(color,q),1/16)            add(2,dest,unit(ach,q),-1/16)
''' + s[b:]a=s.index('# Fourth layer')b=s.index('# Sixth layer',a)s=s[:a]+r'''# Fourth layer: ordinary propagation plus certified dense bridges.for p in range(81):    for ch in (0,1):        carry(3,ch,p)    for color,dch,tch in [(0,2,4),(1,3,5)]:        dest=unit(dch,p)        add(3,dest,unit(dch,p),1)        add(3,dest,unit(tch,p),-4096)        for q in adj[p]:            add(3,dest,unit(color,q),-1)            add(3,dest,unit(dch,q),1)            add(3,dest,unit(tch,q),-4096)        dest=unit(tch,p)        B[3][dest]=-len(adj[p])-1        add(3,dest,unit(color,p),1)        for q in adj[p]:            add(3,dest,unit(color,q),1)        for q in points(p,DI+JU):            add(3,dest,unit(color,q),1/16)            add(3,dest,unit(dch,q),-1/16)            add(3,dest,unit(tch,q),1/16)
# Fifth layer completes long-range propagation and prepares action features.for p in range(81):    for ch in (0,1):        carry(4,ch,p)    for color,dch,tch in [(0,2,4),(1,3,5)]:        dest=unit(dch,p)        add(4,dest,unit(dch,p),1)        add(4,dest,unit(tch,p),-4096)        for q in adj[p]:            add(4,dest,unit(color,q),-1)            add(4,dest,unit(dch,q),1)            add(4,dest,unit(tch,q),-4096)
    # Proactive pressure on enemy stones with few local liberties.    dest=unit(4,p)    B[4][dest]=-len(adj[p])    add(4,dest,unit(1,p),3)    for q in adj[p]:        add(4,dest,unit(0,q),1)        add(4,dest,unit(1,q),0.5)
    dest=unit(5,p)    B[4][dest]=1    for q in adj[p]:        add(4,dest,unit(0,q),-1)        add(4,dest,unit(2,q),1)        add(4,dest,unit(4,q),-4096)
''' + s[b:]s=s.replace('    add(5,dest,unit(4,p),1)','''    for q in adj[p]:        add(5,dest,unit(4,q),P["weak_enemy"])''')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added certified two-step safety propagation through dense strings.")

import json,shutilshutil.copyfile("/work/design.py","/work/design_v10.py")s=open("/work/design_v9.py").read()a=s.index('    # Local liberty pressure')b=s.index('# Fourth layer',a)s=s[:a]+r'''    # Long-range safety messages travel only through proven same-colour    # bridges: a stone and ALL its neighbours must have that colour.    # Every Manhattan-distance-two endpoint is then connected to it.    ring2=points(p,DI+JU)    for color,ach,ch in [(0,2,4),(1,4,5)]:        dest=unit(ch,p)        B[2][dest]=-len(adj[p])-1        add(2,dest,unit(color,p),1)        for q in adj[p]:            add(2,dest,unit(color,q),1)        for q in ring2:            add(2,dest,unit(color,q),1/16)            add(2,dest,unit(ach,q),-1/16)
''' + s[b:]a=s.index('# Fourth layer')b=s.index('# Sixth layer',a)s=s[:a]+r'''# Fourth layer: ordinary propagation plus certified dense bridges.for p in range(81):    for ch in (0,1):        carry(3,ch,p)    for color,dch,tch in [(0,2,4),(1,3,5)]:        dest=unit(dch,p)        add(3,dest,unit(dch,p),1)        add(3,dest,unit(tch,p),-4096)        for q in adj[p]:            add(3,dest,unit(color,q),-1)            add(3,dest,unit(dch,q),1)            add(3,dest,unit(tch,q),-4096)        dest=unit(tch,p)        B[3][dest]=-len(adj[p])-1        add(3,dest,unit(color,p),1)        for q in adj[p]:            add(3,dest,unit(color,q),1)        for q in points(p,DI+JU):            add(3,dest,unit(color,q),1/16)            add(3,dest,unit(dch,q),-1/16)            add(3,dest,unit(tch,q),1/16)
# Fifth layer completes long-range propagation and prepares action features.for p in range(81):    for ch in (0,1):        carry(4,ch,p)    for color,dch,tch in [(0,2,4),(1,3,5)]:        dest=unit(dch,p)        add(4,dest,unit(dch,p),1)        add(4,dest,unit(tch,p),-4096)        for q in adj[p]:            add(4,dest,unit(color,q),-1)            add(4,dest,unit(dch,q),1)            add(4,dest,unit(tch,q),-4096)
    # Proactive pressure on enemy stones with few local liberties.    dest=unit(4,p)    B[4][dest]=-len(adj[p])    add(4,dest,unit(1,p),3)    for q in adj[p]:        add(4,dest,unit(0,q),1)        add(4,dest,unit(1,q),0.5)
    dest=unit(5,p)    B[4][dest]=1    for q in adj[p]:        add(4,dest,unit(0,q),-1)        add(4,dest,unit(2,q),1)        add(4,dest,unit(4,q),-4096)
''' + s[b:]s=s.replace('    add(5,dest,unit(4,p),1)','''    for q in adj[p]:        add(5,dest,unit(4,q),P["weak_enemy"])''')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added certified two-step safety propagation through dense strings.")

OUTPUT of Python run 18 of at most 120:

Games pass 11 of 25 done.

{"ok": true, "pass": 11, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.5416666865348816, "wins": 52, "losses": 44, "score_as_black": 0.375, "score_as_white": 0.7083333134651184, "mean_margin": 36.5625, "mean_length": 113.17708333333333, "stones_captured_per_game": 51.895833333333336, "stones_lost_per_game": 13.916666666666666, "passes_per_game": 2.2083333333333335, "games_with_a_pass_before_move_60": 0, "end_two_passes": 87, "end_repetition": 7, "end_move_limit": 2}, "weak": {"games": 32, "score": 0.9375, "wins": 30, "losses": 2, "score_as_black": 1.0, "score_as_white": 0.875, "mean_margin": 68.40625, "mean_length": 121.65625, "stones_captured_per_game": 80.78125, "stones_lost_per_game": 4.53125, "passes_per_game": 3.375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 29, "end_repetition": 2, "end_move_limit": 1}}, "files": "/work/pass_11/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 11, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.5416666865348816, "wins": 52, "losses": 44, "score_as_black": 0.375, "score_as_white": 0.7083333134651184, "mean_margin": 36.5625, "mean_length": 113.17708333333333, "stones_captured_per_game": 51.895833333333336, "stones_lost_per_game": 13.916666666666666, "passes_per_game": 2.2083333333333335, "games_with_a_pass_before_move_60": 0, "end_two_passes": 87, "end_repetition": 7, "end_move_limit": 2}, "weak": {"games": 32, "score": 0.9375, "wins": 30, "losses": 2, "score_as_black": 1.0, "score_as_white": 0.875, "mean_margin": 68.40625, "mean_length": 121.65625, "stones_captured_per_game": 80.78125, "stones_lost_per_game": 4.53125, "passes_per_game": 3.375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 29, "end_repetition": 2, "end_move_limit": 1}}, "files": "/work/pass_11/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 14.

python

import json,shutilshutil.copyfile("/work/design.py","/work/design_v11.py")s=open("/work/design_v9.py").read()marker='# Tengen is a useful universal first move'idx=s.index(marker)extra=r'''# The unused units certify connectivity along complete straight strings.# A straight occupied segment has no duplicated boundary liberties.segments=[]for r in [0,2,4,6,8]:    segments += [[r*9+c for c in range(5)], [r*9+c for c in range(4,9)]]for c in [0,2,4,6,8]:    segments += [[r*9+c for r in range(5)], [r*9+c for r in range(4,9)]]for r in [1,7]:    segments.append([r*9+c for c in range(2,7)])for c in [1,7]:    segments.append([r*9+c for r in range(2,7)])assert len(segments)==24
# First layer seeds enemy strings with two distinct liberties.for j,seg in enumerate(segments):    u=487+j    boundary=[q for p in seg for q in adj[p] if q not in seg]    assert len(boundary)==len(set(boundary))    B[0][u]=len(boundary)-1-16*len(seg)    for p in seg:        W[0][3*p+1,u]+=16    for q in boundary:        W[0][3*q,u]-=1        W[0][3*q+1,u]-=1    for p in seg:        W[1][u,unit(4,p)]-=16
# Second layer seeds own strings in the same way.for j,seg in enumerate(segments):    u=487+j    boundary=[q for p in seg for q in adj[p] if q not in seg]    B[1][u]=len(boundary)-1-16*len(seg)    for p in seg:        W[1][unit(0,p),u]+=16    for q in boundary:        W[1][unit(0,q),u]-=1        W[1][unit(1,q),u]-=1    reach=set(seg)    for p in seg:        reach.update(adj[p])    for p in reach:        W[2][u,unit(2,p)]-=16
# Further layers broadcast safety across fully connected segments.for k,color,danger in [(2,1,4),(3,0,2)]:    for j,seg in enumerate(segments):        u=487+j        B[k][u]=-len(seg)        for p in seg:            W[k][unit(color,p),u]+=1+1/16            W[k][unit(danger,p),u]-=1/16        reach=set(seg)        for p in seg:            reach.update(adj[p])        outdanger=3 if color==1 else 2        for p in reach:            W[k+1][u,unit(outdanger,p)]-=32
'''s=s[:idx]+extra+s[idx:]open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added exact straight-string liberty and connectivity certificates.")

import json,shutilshutil.copyfile("/work/design.py","/work/design_v11.py")s=open("/work/design_v9.py").read()marker='# Tengen is a useful universal first move'idx=s.index(marker)extra=r'''# The unused units certify connectivity along complete straight strings.# A straight occupied segment has no duplicated boundary liberties.segments=[]for r in [0,2,4,6,8]:    segments += [[r*9+c for c in range(5)], [r*9+c for c in range(4,9)]]for c in [0,2,4,6,8]:    segments += [[r*9+c for r in range(5)], [r*9+c for r in range(4,9)]]for r in [1,7]:    segments.append([r*9+c for c in range(2,7)])for c in [1,7]:    segments.append([r*9+c for r in range(2,7)])assert len(segments)==24
# First layer seeds enemy strings with two distinct liberties.for j,seg in enumerate(segments):    u=487+j    boundary=[q for p in seg for q in adj[p] if q not in seg]    assert len(boundary)==len(set(boundary))    B[0][u]=len(boundary)-1-16*len(seg)    for p in seg:        W[0][3*p+1,u]+=16    for q in boundary:        W[0][3*q,u]-=1        W[0][3*q+1,u]-=1    for p in seg:        W[1][u,unit(4,p)]-=16
# Second layer seeds own strings in the same way.for j,seg in enumerate(segments):    u=487+j    boundary=[q for p in seg for q in adj[p] if q not in seg]    B[1][u]=len(boundary)-1-16*len(seg)    for p in seg:        W[1][unit(0,p),u]+=16    for q in boundary:        W[1][unit(0,q),u]-=1        W[1][unit(1,q),u]-=1    reach=set(seg)    for p in seg:        reach.update(adj[p])    for p in reach:        W[2][u,unit(2,p)]-=16
# Further layers broadcast safety across fully connected segments.for k,color,danger in [(2,1,4),(3,0,2)]:    for j,seg in enumerate(segments):        u=487+j        B[k][u]=-len(seg)        for p in seg:            W[k][unit(color,p),u]+=1+1/16            W[k][unit(danger,p),u]-=1/16        reach=set(seg)        for p in seg:            reach.update(adj[p])        outdanger=3 if color==1 else 2        for p in reach:            W[k+1][u,unit(outdanger,p)]-=32
'''s=s[:idx]+extra+s[idx:]open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added exact straight-string liberty and connectivity certificates.")

OUTPUT of Python run 19 of at most 120:

Games pass 12 of 25 done.

{"ok": true, "pass": 12, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.6979166865348816, "wins": 67, "losses": 29, "score_as_black": 0.5833333134651184, "score_as_white": 0.8125, "mean_margin": 44.57291793823242, "mean_length": 119.48958333333333, "stones_captured_per_game": 67.94791666666667, "stones_lost_per_game": 13.427083333333334, "passes_per_game": 2.4583333333333335, "games_with_a_pass_before_move_60": 0, "end_two_passes": 86, "end_repetition": 6, "end_move_limit": 4}, "weak": {"games": 32, "score": 0.90625, "wins": 29, "losses": 3, "score_as_black": 1.0, "score_as_white": 0.8125, "mean_margin": 71.5625, "mean_length": 122.15625, "stones_captured_per_game": 88.125, "stones_lost_per_game": 2.84375, "passes_per_game": 2.9375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 29, "end_repetition": 3, "end_move_limit": 0}}, "files": "/work/pass_12/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 12, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.6979166865348816, "wins": 67, "losses": 29, "score_as_black": 0.5833333134651184, "score_as_white": 0.8125, "mean_margin": 44.57291793823242, "mean_length": 119.48958333333333, "stones_captured_per_game": 67.94791666666667, "stones_lost_per_game": 13.427083333333334, "passes_per_game": 2.4583333333333335, "games_with_a_pass_before_move_60": 0, "end_two_passes": 86, "end_repetition": 6, "end_move_limit": 4}, "weak": {"games": 32, "score": 0.90625, "wins": 29, "losses": 3, "score_as_black": 1.0, "score_as_white": 0.8125, "mean_margin": 71.5625, "mean_length": 122.15625, "stones_captured_per_game": 88.125, "stones_lost_per_game": 2.84375, "passes_per_game": 2.9375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 29, "end_repetition": 3, "end_move_limit": 0}}, "files": "/work/pass_12/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 13.

python

import json,shutilshutil.copyfile("/work/design.py","/work/design_v12.py")s=open("/work/design.py").read()idx=s.index('# Tengen is a useful universal first move')extra=r'''# Corner/edge snapback and cyclic-capture guard.# A singleton capture is dangerous if the resulting string has no# other empty neighbour and no connection to a healthy friendly string.corner_points=set()for r0,c0,dr,dc in [(0,0,1,1),(0,8,1,-1),(8,0,-1,1),(8,8,-1,-1)]:    for a,b in [(0,0),(1,0),(0,1),(1,1),(2,0),(0,2)]:        corner_points.add((r0+dr*a)*9+c0+dc*b)corner_points=sorted(corner_points)assert len(corner_points)==24slot={p:487+j for j,p in enumerate(corner_points)}for p,u in slot.items():    W[4][unit(3,p),u]=1    for q in adj[p]:        W[4][unit(1,q),u]-=1
for p,u in slot.items():    B[5][u]=-4-4*len(adj[p])    W[5][unit(5,p),u]+=4    for q in adj[p]:        W[5][unit(0,q),u]+=4        W[5][unit(1,q),u]+=4        W[5][unit(3,q),u]-=2        if q in slot:            W[5][slot[q],u]+=3    OW[u,p]-=14
'''s=s[:idx]+extra+s[idx:]open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added local snapback/cyclic-capture guards in all four corners.")

import json,shutilshutil.copyfile("/work/design.py","/work/design_v12.py")s=open("/work/design.py").read()idx=s.index('# Tengen is a useful universal first move')extra=r'''# Corner/edge snapback and cyclic-capture guard.# A singleton capture is dangerous if the resulting string has no# other empty neighbour and no connection to a healthy friendly string.corner_points=set()for r0,c0,dr,dc in [(0,0,1,1),(0,8,1,-1),(8,0,-1,1),(8,8,-1,-1)]:    for a,b in [(0,0),(1,0),(0,1),(1,1),(2,0),(0,2)]:        corner_points.add((r0+dr*a)*9+c0+dc*b)corner_points=sorted(corner_points)assert len(corner_points)==24slot={p:487+j for j,p in enumerate(corner_points)}for p,u in slot.items():    W[4][unit(3,p),u]=1    for q in adj[p]:        W[4][unit(1,q),u]-=1
for p,u in slot.items():    B[5][u]=-4-4*len(adj[p])    W[5][unit(5,p),u]+=4    for q in adj[p]:        W[5][unit(0,q),u]+=4        W[5][unit(1,q),u]+=4        W[5][unit(3,q),u]-=2        if q in slot:            W[5][slot[q],u]+=3    OW[u,p]-=14
'''s=s[:idx]+extra+s[idx:]open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added local snapback/cyclic-capture guards in all four corners.")

OUTPUT of Python run 20 of at most 120:

Games pass 13 of 25 done.

{"ok": true, "pass": 13, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.5208333134651184, "wins": 50, "losses": 46, "score_as_black": 0.3958333432674408, "score_as_white": 0.6458333134651184, "mean_margin": 30.36458396911621, "mean_length": 111.69791666666667, "stones_captured_per_game": 49.677083333333336, "stones_lost_per_game": 15.114583333333334, "passes_per_game": 2.1354166666666665, "games_with_a_pass_before_move_60": 0, "end_two_passes": 88, "end_repetition": 6, "end_move_limit": 2}, "weak": {"games": 32, "score": 0.96875, "wins": 31, "losses": 1, "score_as_black": 0.9375, "score_as_white": 1.0, "mean_margin": 70.75, "mean_length": 121.40625, "stones_captured_per_game": 79.6875, "stones_lost_per_game": 3.03125, "passes_per_game": 2.84375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 30, "end_repetition": 2, "end_move_limit": 0}}, "files": "/work/pass_13/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 13, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.5208333134651184, "wins": 50, "losses": 46, "score_as_black": 0.3958333432674408, "score_as_white": 0.6458333134651184, "mean_margin": 30.36458396911621, "mean_length": 111.69791666666667, "stones_captured_per_game": 49.677083333333336, "stones_lost_per_game": 15.114583333333334, "passes_per_game": 2.1354166666666665, "games_with_a_pass_before_move_60": 0, "end_two_passes": 88, "end_repetition": 6, "end_move_limit": 2}, "weak": {"games": 32, "score": 0.96875, "wins": 31, "losses": 1, "score_as_black": 0.9375, "score_as_white": 1.0, "mean_margin": 70.75, "mean_length": 121.40625, "stones_captured_per_game": 79.6875, "stones_lost_per_game": 3.03125, "passes_per_game": 2.84375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 30, "end_repetition": 2, "end_move_limit": 0}}, "files": "/work/pass_13/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 12.

python

import json,shutilshutil.copyfile("/work/design.py","/work/design_v13.py")s=open("/work/design_v12.py").read()s=s.replace('for k,color,danger in [(2,1,4),(3,0,2)]:',            'for k,color,danger in [(2,1,4),(3,1,3)]:')needle='        reach=set(seg)\n        for p in seg:\n            reach.update(adj[p])\n        outdanger=3 if color==1 else 2'replacement='''        if k==3:            # Intersecting certified segments belong to the same string.            for jj,other in enumerate(segments):                if set(seg).intersection(other):                    W[k][487+jj,u]+=1/128        reach=set(seg)        for p in seg:            reach.update(adj[p])        outdanger=3 if color==1 else 2'''assert needle in ss=s.replace(needle,replacement)# A message can now be as small as (1/16)/128.s=s.replace('W[k+1][u,unit(outdanger,p)]-=32',            'W[k+1][u,unit(outdanger,p)]-=4096')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Extended enemy-string safety through intersecting certified segments.")

import json,shutilshutil.copyfile("/work/design.py","/work/design_v13.py")s=open("/work/design_v12.py").read()s=s.replace('for k,color,danger in [(2,1,4),(3,0,2)]:',            'for k,color,danger in [(2,1,4),(3,1,3)]:')needle='        reach=set(seg)\n        for p in seg:\n            reach.update(adj[p])\n        outdanger=3 if color==1 else 2'replacement='''        if k==3:            # Intersecting certified segments belong to the same string.            for jj,other in enumerate(segments):                if set(seg).intersection(other):                    W[k][487+jj,u]+=1/128        reach=set(seg)        for p in seg:            reach.update(adj[p])        outdanger=3 if color==1 else 2'''assert needle in ss=s.replace(needle,replacement)# A message can now be as small as (1/16)/128.s=s.replace('W[k+1][u,unit(outdanger,p)]-=32',            'W[k+1][u,unit(outdanger,p)]-=4096')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Extended enemy-string safety through intersecting certified segments.")

OUTPUT of Python run 21 of at most 120:

Games pass 14 of 25 done.

{"ok": true, "pass": 14, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.6354166865348816, "wins": 61, "losses": 35, "score_as_black": 0.5, "score_as_white": 0.7708333134651184, "mean_margin": 39.57291793823242, "mean_length": 114.27083333333333, "stones_captured_per_game": 57.135416666666664, "stones_lost_per_game": 12.489583333333334, "passes_per_game": 1.9895833333333333, "games_with_a_pass_before_move_60": 0, "end_two_passes": 89, "end_repetition": 5, "end_move_limit": 2}, "weak": {"games": 32, "score": 0.96875, "wins": 31, "losses": 1, "score_as_black": 0.9375, "score_as_white": 1.0, "mean_margin": 71.75, "mean_length": 127.09375, "stones_captured_per_game": 84.40625, "stones_lost_per_game": 4.71875, "passes_per_game": 3.71875, "games_with_a_pass_before_move_60": 0, "end_two_passes": 30, "end_repetition": 1, "end_move_limit": 1}}, "files": "/work/pass_14/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 14, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.6354166865348816, "wins": 61, "losses": 35, "score_as_black": 0.5, "score_as_white": 0.7708333134651184, "mean_margin": 39.57291793823242, "mean_length": 114.27083333333333, "stones_captured_per_game": 57.135416666666664, "stones_lost_per_game": 12.489583333333334, "passes_per_game": 1.9895833333333333, "games_with_a_pass_before_move_60": 0, "end_two_passes": 89, "end_repetition": 5, "end_move_limit": 2}, "weak": {"games": 32, "score": 0.96875, "wins": 31, "losses": 1, "score_as_black": 0.9375, "score_as_white": 1.0, "mean_margin": 71.75, "mean_length": 127.09375, "stones_captured_per_game": 84.40625, "stones_lost_per_game": 4.71875, "passes_per_game": 3.71875, "games_with_a_pass_before_move_60": 0, "end_two_passes": 30, "end_repetition": 1, "end_move_limit": 1}}, "files": "/work/pass_14/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 11.

python

import json,shutilshutil.copyfile("/work/design.py","/work/design_v14.py")overrides=json.load(open("/work/params_v9.json"))overrides["decisiveness"]=2.5open("/work/params_v15.json","w").write(json.dumps(overrides))exec(open("/work/design.py").read())print("Made the certified tactical/shape policy more decisive.")

import json,shutilshutil.copyfile("/work/design.py","/work/design_v14.py")overrides=json.load(open("/work/params_v9.json"))overrides["decisiveness"]=2.5open("/work/params_v15.json","w").write(json.dumps(overrides))exec(open("/work/design.py").read())print("Made the certified tactical/shape policy more decisive.")

OUTPUT of Python run 22 of at most 120:

Games pass 15 of 25 done.

{"ok": true, "pass": 15, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.59375, "wins": 57, "losses": 39, "score_as_black": 0.4791666567325592, "score_as_white": 0.7083333134651184, "mean_margin": 37.63541793823242, "mean_length": 115.14583333333333, "stones_captured_per_game": 56.40625, "stones_lost_per_game": 13.8125, "passes_per_game": 2.1458333333333335, "games_with_a_pass_before_move_60": 0, "end_two_passes": 89, "end_repetition": 4, "end_move_limit": 3}, "weak": {"games": 32, "score": 0.96875, "wins": 31, "losses": 1, "score_as_black": 1.0, "score_as_white": 0.9375, "mean_margin": 78.6875, "mean_length": 125.46875, "stones_captured_per_game": 93.8125, "stones_lost_per_game": 2.625, "passes_per_game": 2.875, "games_with_a_pass_before_move_60": 0, "end_two_passes": 30, "end_repetition": 2, "end_move_limit": 0}}, "files": "/work/pass_15/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 15, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.59375, "wins": 57, "losses": 39, "score_as_black": 0.4791666567325592, "score_as_white": 0.7083333134651184, "mean_margin": 37.63541793823242, "mean_length": 115.14583333333333, "stones_captured_per_game": 56.40625, "stones_lost_per_game": 13.8125, "passes_per_game": 2.1458333333333335, "games_with_a_pass_before_move_60": 0, "end_two_passes": 89, "end_repetition": 4, "end_move_limit": 3}, "weak": {"games": 32, "score": 0.96875, "wins": 31, "losses": 1, "score_as_black": 1.0, "score_as_white": 0.9375, "mean_margin": 78.6875, "mean_length": 125.46875, "stones_captured_per_game": 93.8125, "stones_lost_per_game": 2.625, "passes_per_game": 2.875, "games_with_a_pass_before_move_60": 0, "end_two_passes": 30, "end_repetition": 2, "end_move_limit": 0}}, "files": "/work/pass_15/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 10.

python

import json,shutilshutil.copyfile("/work/design.py","/work/design_v15.py")s=open("/work/design_v14.py").read()s=s.replace('        add(2,dest,unit(color,p),3)\n        B[2][dest]=-len(adj[p])','''        add(2,dest,unit(color,p),7)        B[2][dest]=-len(adj[p])-4''')s=s.replace('            add(2,dest,unit(color,q),-0.5)','''            add(2,dest,unit(color,q),-0.5)            # Only healthy neighbours reduce a stone's local urgency.            add(2,dest,unit(2 if color==0 else 4,q),0.5)''')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Local urgency now discounts healthy support, not endangered string-mates.")

import json,shutilshutil.copyfile("/work/design.py","/work/design_v15.py")s=open("/work/design_v14.py").read()s=s.replace('        add(2,dest,unit(color,p),3)\n        B[2][dest]=-len(adj[p])','''        add(2,dest,unit(color,p),7)        B[2][dest]=-len(adj[p])-4''')s=s.replace('            add(2,dest,unit(color,q),-0.5)','''            add(2,dest,unit(color,q),-0.5)            # Only healthy neighbours reduce a stone's local urgency.            add(2,dest,unit(2 if color==0 else 4,q),0.5)''')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Local urgency now discounts healthy support, not endangered string-mates.")

OUTPUT of Python run 23 of at most 120:

Games pass 16 of 25 done.

{"ok": true, "pass": 16, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.625, "wins": 60, "losses": 36, "score_as_black": 0.5416666865348816, "score_as_white": 0.7083333134651184, "mean_margin": 42.45833206176758, "mean_length": 115.86458333333333, "stones_captured_per_game": 60.65625, "stones_lost_per_game": 11.947916666666666, "passes_per_game": 1.8958333333333333, "games_with_a_pass_before_move_60": 0, "end_two_passes": 80, "end_repetition": 12, "end_move_limit": 4}, "weak": {"games": 32, "score": 0.90625, "wins": 29, "losses": 3, "score_as_black": 0.875, "score_as_white": 0.9375, "mean_margin": 71.1875, "mean_length": 121.0625, "stones_captured_per_game": 81.09375, "stones_lost_per_game": 4.71875, "passes_per_game": 2.28125, "games_with_a_pass_before_move_60": 0, "end_two_passes": 31, "end_repetition": 1, "end_move_limit": 0}}, "files": "/work/pass_16/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 16, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.625, "wins": 60, "losses": 36, "score_as_black": 0.5416666865348816, "score_as_white": 0.7083333134651184, "mean_margin": 42.45833206176758, "mean_length": 115.86458333333333, "stones_captured_per_game": 60.65625, "stones_lost_per_game": 11.947916666666666, "passes_per_game": 1.8958333333333333, "games_with_a_pass_before_move_60": 0, "end_two_passes": 80, "end_repetition": 12, "end_move_limit": 4}, "weak": {"games": 32, "score": 0.90625, "wins": 29, "losses": 3, "score_as_black": 0.875, "score_as_white": 0.9375, "mean_margin": 71.1875, "mean_length": 121.0625, "stones_captured_per_game": 81.09375, "stones_lost_per_game": 4.71875, "passes_per_game": 2.28125, "games_with_a_pass_before_move_60": 0, "end_two_passes": 31, "end_repetition": 1, "end_move_limit": 0}}, "files": "/work/pass_16/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 9.

python

import json,shutilshutil.copyfile("/work/design.py","/work/design_v16.py")s=open("/work/design.py").read()idx=s.index('# Tengen is a useful universal first move')extra=r'''# Conservative endgame stopping: count only stones certified to have# at least two liberties. Ignore all prospective territory and captures.# If those stones alone beat every other point plus komi, avoid needless# invasions and positional-repetition fights.W[0][2,511]=1for k in range(1,6):    W[k][511,511]=1for u,shift in [(487,0),(488,1)]:    B[5][u]=-46.75-shift    W[5][511,u]=7.5    for p in range(81):        W[5][unit(0,p),u]+=1        W[5][unit(2,p),u]-=1OW[487,81]+=16OW[488,81]-=16
'''s=s[:idx]+extra+s[idx:]open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added a conservative safe-stone, komi-aware winning-pass gate.")

import json,shutilshutil.copyfile("/work/design.py","/work/design_v16.py")s=open("/work/design.py").read()idx=s.index('# Tengen is a useful universal first move')extra=r'''# Conservative endgame stopping: count only stones certified to have# at least two liberties. Ignore all prospective territory and captures.# If those stones alone beat every other point plus komi, avoid needless# invasions and positional-repetition fights.W[0][2,511]=1for k in range(1,6):    W[k][511,511]=1for u,shift in [(487,0),(488,1)]:    B[5][u]=-46.75-shift    W[5][511,u]=7.5    for p in range(81):        W[5][unit(0,p),u]+=1        W[5][unit(2,p),u]-=1OW[487,81]+=16OW[488,81]-=16
'''s=s[:idx]+extra+s[idx:]open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added a conservative safe-stone, komi-aware winning-pass gate.")

OUTPUT of Python run 24 of at most 120:

Games pass 17 of 25 done.

{"ok": true, "pass": 17, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.5729166865348816, "wins": 55, "losses": 41, "score_as_black": 0.5416666865348816, "score_as_white": 0.6041666865348816, "mean_margin": 26.76041603088379, "mean_length": 123.40625, "stones_captured_per_game": 55.53125, "stones_lost_per_game": 15.333333333333334, "passes_per_game": 8.770833333333334, "games_with_a_pass_before_move_60": 0, "end_two_passes": 69, "end_repetition": 6, "end_move_limit": 21}, "weak": {"games": 32, "score": 0.84375, "wins": 27, "losses": 5, "score_as_black": 0.8125, "score_as_white": 0.875, "mean_margin": 53.03125, "mean_length": 140.8125, "stones_captured_per_game": 92.71875, "stones_lost_per_game": 6.71875, "passes_per_game": 17.0, "games_with_a_pass_before_move_60": 0, "end_two_passes": 15, "end_repetition": 4, "end_move_limit": 13}}, "files": "/work/pass_17/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 17, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.5729166865348816, "wins": 55, "losses": 41, "score_as_black": 0.5416666865348816, "score_as_white": 0.6041666865348816, "mean_margin": 26.76041603088379, "mean_length": 123.40625, "stones_captured_per_game": 55.53125, "stones_lost_per_game": 15.333333333333334, "passes_per_game": 8.770833333333334, "games_with_a_pass_before_move_60": 0, "end_two_passes": 69, "end_repetition": 6, "end_move_limit": 21}, "weak": {"games": 32, "score": 0.84375, "wins": 27, "losses": 5, "score_as_black": 0.8125, "score_as_white": 0.875, "mean_margin": 53.03125, "mean_length": 140.8125, "stones_captured_per_game": 92.71875, "stones_lost_per_game": 6.71875, "passes_per_game": 17.0, "games_with_a_pass_before_move_60": 0, "end_two_passes": 15, "end_repetition": 4, "end_move_limit": 13}}, "files": "/work/pass_17/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 8.

python

import json,shutilshutil.copyfile("/work/design.py","/work/design_v17.py")s=open("/work/design_v14.py").read()a=s.index('    # Local liberty pressure')b=s.index('# Fourth layer',a)s=s[:a]+r'''    # Shared friendly/enemy contact supports blocking, hane and connection.    # min(own-contact, enemy-contact) is reconstructed in the next layer.    dest=unit(4,p)    for q in adj[p]:        add(2,dest,unit(0,q),1)        add(2,dest,unit(1,q),-1)    dest=unit(5,p)    add(2,dest,unit(1,p),3)    B[2][dest]=-len(adj[p])    for q in adj[p]:        add(2,dest,unit(0,q),1)        add(2,dest,unit(1,q),0.5)
''' + s[b:]a=s.index('    # Urgency to support')b=s.index('    # Binary access',a)s=s[:a]+r'''    B[3][unit(4,p)]=20    add(3,unit(4,p),unit(4,p),-1.5)    for q in adj[p]:        add(3,unit(4,p),unit(0,q),1.5)        add(3,unit(4,p),unit(5,q),P["weak_enemy"])
''' + s[b:]s=s.replace('    B[5][dest] = 20','    B[5][dest] = 0')# The carried pressure/contact channel already contains the +20 offset.open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added explicit mixed-contact blocking/hane/connection preference.")

import json,shutilshutil.copyfile("/work/design.py","/work/design_v17.py")s=open("/work/design_v14.py").read()a=s.index('    # Local liberty pressure')b=s.index('# Fourth layer',a)s=s[:a]+r'''    # Shared friendly/enemy contact supports blocking, hane and connection.    # min(own-contact, enemy-contact) is reconstructed in the next layer.    dest=unit(4,p)    for q in adj[p]:        add(2,dest,unit(0,q),1)        add(2,dest,unit(1,q),-1)    dest=unit(5,p)    add(2,dest,unit(1,p),3)    B[2][dest]=-len(adj[p])    for q in adj[p]:        add(2,dest,unit(0,q),1)        add(2,dest,unit(1,q),0.5)
''' + s[b:]a=s.index('    # Urgency to support')b=s.index('    # Binary access',a)s=s[:a]+r'''    B[3][unit(4,p)]=20    add(3,unit(4,p),unit(4,p),-1.5)    for q in adj[p]:        add(3,unit(4,p),unit(0,q),1.5)        add(3,unit(4,p),unit(5,q),P["weak_enemy"])
''' + s[b:]s=s.replace('    B[5][dest] = 20','    B[5][dest] = 0')# The carried pressure/contact channel already contains the +20 offset.open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added explicit mixed-contact blocking/hane/connection preference.")

OUTPUT of Python run 25 of at most 120:

Games pass 18 of 25 done.

{"ok": true, "pass": 18, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.6666666865348816, "wins": 64, "losses": 32, "score_as_black": 0.5833333134651184, "score_as_white": 0.75, "mean_margin": 43.45833206176758, "mean_length": 114.8125, "stones_captured_per_game": 55.489583333333336, "stones_lost_per_game": 14.041666666666666, "passes_per_game": 1.59375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 84, "end_repetition": 11, "end_move_limit": 1}, "weak": {"games": 32, "score": 1.0, "wins": 32, "losses": 0, "score_as_black": 1.0, "score_as_white": 1.0, "mean_margin": 76.40625, "mean_length": 121.9375, "stones_captured_per_game": 85.15625, "stones_lost_per_game": 3.5, "passes_per_game": 2.40625, "games_with_a_pass_before_move_60": 0, "end_two_passes": 32, "end_repetition": 0, "end_move_limit": 0}}, "files": "/work/pass_18/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 18, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.6666666865348816, "wins": 64, "losses": 32, "score_as_black": 0.5833333134651184, "score_as_white": 0.75, "mean_margin": 43.45833206176758, "mean_length": 114.8125, "stones_captured_per_game": 55.489583333333336, "stones_lost_per_game": 14.041666666666666, "passes_per_game": 1.59375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 84, "end_repetition": 11, "end_move_limit": 1}, "weak": {"games": 32, "score": 1.0, "wins": 32, "losses": 0, "score_as_black": 1.0, "score_as_white": 1.0, "mean_margin": 76.40625, "mean_length": 121.9375, "stones_captured_per_game": 85.15625, "stones_lost_per_game": 3.5, "passes_per_game": 2.40625, "games_with_a_pass_before_move_60": 0, "end_two_passes": 32, "end_repetition": 0, "end_move_limit": 0}}, "files": "/work/pass_18/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 7.

python

import json,shutilshutil.copyfile("/work/design.py","/work/design_v18.py")s=open("/work/design.py").read()a=s.index('segments=[]')b=s.index('assert len(segments)==24',a)s=s[:a]+r'''# Fully occupied 2x2 blocks are connected and have eight distinct# external neighbours, so their liberty certificates are exact too.anchors=[(r,c) for r in [0,2,5,7] for c in [0,2,5,7]]anchors += [(r,c) for r in [3,4] for c in [3,4]]anchors += [(1,1),(1,6),(6,1),(6,6)]segments=[[(r+dr)*9+c+dc for dr in [0,1] for dc in [0,1]]          for r,c in anchors]''' +s[b:]open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Connectivity certificates now cover dense 2x2 blocks across the board.")

import json,shutilshutil.copyfile("/work/design.py","/work/design_v18.py")s=open("/work/design.py").read()a=s.index('segments=[]')b=s.index('assert len(segments)==24',a)s=s[:a]+r'''# Fully occupied 2x2 blocks are connected and have eight distinct# external neighbours, so their liberty certificates are exact too.anchors=[(r,c) for r in [0,2,5,7] for c in [0,2,5,7]]anchors += [(r,c) for r in [3,4] for c in [3,4]]anchors += [(1,1),(1,6),(6,1),(6,6)]segments=[[(r+dr)*9+c+dc for dr in [0,1] for dc in [0,1]]          for r,c in anchors]''' +s[b:]open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Connectivity certificates now cover dense 2x2 blocks across the board.")

OUTPUT of Python run 26 of at most 120:

Games pass 19 of 25 done.

{"ok": true, "pass": 19, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.7604166865348816, "wins": 73, "losses": 23, "score_as_black": 0.7083333134651184, "score_as_white": 0.8125, "mean_margin": 55.46875, "mean_length": 119.35416666666667, "stones_captured_per_game": 68.23958333333333, "stones_lost_per_game": 10.791666666666666, "passes_per_game": 1.7604166666666667, "games_with_a_pass_before_move_60": 0, "end_two_passes": 86, "end_repetition": 9, "end_move_limit": 1}, "weak": {"games": 32, "score": 0.96875, "wins": 31, "losses": 1, "score_as_black": 1.0, "score_as_white": 0.9375, "mean_margin": 76.78125, "mean_length": 120.875, "stones_captured_per_game": 83.25, "stones_lost_per_game": 3.0625, "passes_per_game": 2.25, "games_with_a_pass_before_move_60": 0, "end_two_passes": 30, "end_repetition": 2, "end_move_limit": 0}}, "files": "/work/pass_19/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 19, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.7604166865348816, "wins": 73, "losses": 23, "score_as_black": 0.7083333134651184, "score_as_white": 0.8125, "mean_margin": 55.46875, "mean_length": 119.35416666666667, "stones_captured_per_game": 68.23958333333333, "stones_lost_per_game": 10.791666666666666, "passes_per_game": 1.7604166666666667, "games_with_a_pass_before_move_60": 0, "end_two_passes": 86, "end_repetition": 9, "end_move_limit": 1}, "weak": {"games": 32, "score": 0.96875, "wins": 31, "losses": 1, "score_as_black": 1.0, "score_as_white": 0.9375, "mean_margin": 76.78125, "mean_length": 120.875, "stones_captured_per_game": 83.25, "stones_lost_per_game": 3.0625, "passes_per_game": 2.25, "games_with_a_pass_before_move_60": 0, "end_two_passes": 30, "end_repetition": 2, "end_move_limit": 0}}, "files": "/work/pass_19/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 6.

python

import json,shutilshutil.copyfile("/work/design.py","/work/design_v19.py")s=open("/work/design.py").read()needle='# Further layers broadcast safety across fully connected segments.\nfor k,color,danger in [(2,1,4),(3,1,3)]:\n    for j,seg in enumerate(segments):'replacement=r'''# Stagger the later block certificates. Together the three enemy# stages cover all 64 possible 2x2 blocks, preserving board symmetries.def blocks_with_orbits(orbits):    out=[]    for r in range(8):        for c in range(8):            key=tuple(sorted([abs(2*r-7),abs(2*c-7)]))            if key in orbits:                out.append([(r+dr)*9+c+dc for dr in [0,1] for dc in [0,1]])    return outthird=blocks_with_orbits({(1,3),(5,7),(1,1),(5,5)})fourth=blocks_with_orbits({(1,5),(1,7),(3,5)})assert len(third)==len(fourth)==24assert len({tuple(x) for x in segments+third+fourth})==64
# Further layers broadcast safety across fully connected blocks.for k,color,danger in [(2,1,4),(3,1,3)]:    active_segments=third if k==2 else fourth    for j,seg in enumerate(active_segments):'''assert needle in ss=s.replace(needle,replacement)s=s.replace('for jj,other in enumerate(segments):','for jj,other in enumerate(third):')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Staggered safety certificates cover every 2x2 block.")

import json,shutilshutil.copyfile("/work/design.py","/work/design_v19.py")s=open("/work/design.py").read()needle='# Further layers broadcast safety across fully connected segments.\nfor k,color,danger in [(2,1,4),(3,1,3)]:\n    for j,seg in enumerate(segments):'replacement=r'''# Stagger the later block certificates. Together the three enemy# stages cover all 64 possible 2x2 blocks, preserving board symmetries.def blocks_with_orbits(orbits):    out=[]    for r in range(8):        for c in range(8):            key=tuple(sorted([abs(2*r-7),abs(2*c-7)]))            if key in orbits:                out.append([(r+dr)*9+c+dc for dr in [0,1] for dc in [0,1]])    return outthird=blocks_with_orbits({(1,3),(5,7),(1,1),(5,5)})fourth=blocks_with_orbits({(1,5),(1,7),(3,5)})assert len(third)==len(fourth)==24assert len({tuple(x) for x in segments+third+fourth})==64
# Further layers broadcast safety across fully connected blocks.for k,color,danger in [(2,1,4),(3,1,3)]:    active_segments=third if k==2 else fourth    for j,seg in enumerate(active_segments):'''assert needle in ss=s.replace(needle,replacement)s=s.replace('for jj,other in enumerate(segments):','for jj,other in enumerate(third):')open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Staggered safety certificates cover every 2x2 block.")

OUTPUT of Python run 27 of at most 120:

Games pass 20 of 25 done.

{"ok": true, "pass": 20, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.7083333134651184, "wins": 68, "losses": 28, "score_as_black": 0.7291666865348816, "score_as_white": 0.6875, "mean_margin": 46.0625, "mean_length": 116.375, "stones_captured_per_game": 61.25, "stones_lost_per_game": 14.572916666666666, "passes_per_game": 1.6979166666666667, "games_with_a_pass_before_move_60": 1, "end_two_passes": 82, "end_repetition": 10, "end_move_limit": 4}, "weak": {"games": 32, "score": 0.90625, "wins": 29, "losses": 3, "score_as_black": 0.9375, "score_as_white": 0.875, "mean_margin": 69.875, "mean_length": 115.40625, "stones_captured_per_game": 69.09375, "stones_lost_per_game": 2.375, "passes_per_game": 2.09375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 28, "end_repetition": 4, "end_move_limit": 0}}, "files": "/work/pass_20/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 20, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.7083333134651184, "wins": 68, "losses": 28, "score_as_black": 0.7291666865348816, "score_as_white": 0.6875, "mean_margin": 46.0625, "mean_length": 116.375, "stones_captured_per_game": 61.25, "stones_lost_per_game": 14.572916666666666, "passes_per_game": 1.6979166666666667, "games_with_a_pass_before_move_60": 1, "end_two_passes": 82, "end_repetition": 10, "end_move_limit": 4}, "weak": {"games": 32, "score": 0.90625, "wins": 29, "losses": 3, "score_as_black": 0.9375, "score_as_white": 0.875, "mean_margin": 69.875, "mean_length": 115.40625, "stones_captured_per_game": 69.09375, "stones_lost_per_game": 2.375, "passes_per_game": 2.09375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 28, "end_repetition": 4, "end_move_limit": 0}}, "files": "/work/pass_20/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 5.

python

import json,shutilshutil.copyfile("/work/design.py","/work/design_v20.py")s=open("/work/design_v19.py").read()# Insert after the ordinary opening gate has been constructed.idx=s.index('OB[81] = P["pass_bias"]')extra=r'''# Stop unnecessary fights only with BOTH a conservative safe-stone# area lead and two solid eyes witnessed on a symmetric sample.W[0][2,511]=1for k in range(1,4):    W[k][511,511]=1
eye_points=[r*9+c for r in range(0,9,2) for c in range(0,9,2)            if (r,c)!=(4,4)]assert len(eye_points)==24for j,p in enumerate(eye_points):    u=487+j    ring=adj[p]+diag[p]    B[4][u]=1-len(ring)    for q in ring:        W[4][unit(0,q),u]+=1    W[4][unit(0,p),u]-=8    W[4][unit(1,p),u]-=8
# Two neighbouring thresholds make a clipped, komi-aware lead signal.W[4][:,486]=0W[4][:,511]=0for u,shift in [(511,0),(486,1)]:    B[4][u]=-46.75-shift    W[4][511,u]=7.5    for p in range(81):        W[4][unit(0,p),u]+=1        W[4][unit(2,p),u]-=1
# For fewer than two witnessed eyes both gates are zero.# Otherwise their difference is exactly the clipped lead signal.for u,source in [(487,511),(488,486)]:    B[5][u]=-200    W[5][source,u]=1    for v in range(487,511):        W[5][v,u]=100OW[487,81]+=16OW[488,81]-=16
# Recompute the empty-board gate directly; its old carry slot was reused.W[5][:,486]=0B[5][486]=1for p in range(81):    W[5][unit(0,p),486]=-1    W[5][unit(1,p),486]=-1
'''s=s[:idx]+extra+s[idx:]open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added a bounded winning-pass gate requiring two solid eyes and an area lead.")

import json,shutilshutil.copyfile("/work/design.py","/work/design_v20.py")s=open("/work/design_v19.py").read()# Insert after the ordinary opening gate has been constructed.idx=s.index('OB[81] = P["pass_bias"]')extra=r'''# Stop unnecessary fights only with BOTH a conservative safe-stone# area lead and two solid eyes witnessed on a symmetric sample.W[0][2,511]=1for k in range(1,4):    W[k][511,511]=1
eye_points=[r*9+c for r in range(0,9,2) for c in range(0,9,2)            if (r,c)!=(4,4)]assert len(eye_points)==24for j,p in enumerate(eye_points):    u=487+j    ring=adj[p]+diag[p]    B[4][u]=1-len(ring)    for q in ring:        W[4][unit(0,q),u]+=1    W[4][unit(0,p),u]-=8    W[4][unit(1,p),u]-=8
# Two neighbouring thresholds make a clipped, komi-aware lead signal.W[4][:,486]=0W[4][:,511]=0for u,shift in [(511,0),(486,1)]:    B[4][u]=-46.75-shift    W[4][511,u]=7.5    for p in range(81):        W[4][unit(0,p),u]+=1        W[4][unit(2,p),u]-=1
# For fewer than two witnessed eyes both gates are zero.# Otherwise their difference is exactly the clipped lead signal.for u,source in [(487,511),(488,486)]:    B[5][u]=-200    W[5][source,u]=1    for v in range(487,511):        W[5][v,u]=100OW[487,81]+=16OW[488,81]-=16
# Recompute the empty-board gate directly; its old carry slot was reused.W[5][:,486]=0B[5][486]=1for p in range(81):    W[5][unit(0,p),486]=-1    W[5][unit(1,p),486]=-1
'''s=s[:idx]+extra+s[idx:]open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added a bounded winning-pass gate requiring two solid eyes and an area lead.")

OUTPUT of Python run 28 of at most 120:

Games pass 21 of 25 done.

{"ok": true, "pass": 21, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.75, "wins": 72, "losses": 24, "score_as_black": 0.7083333134651184, "score_as_white": 0.7916666865348816, "mean_margin": 48.375, "mean_length": 117.54166666666667, "stones_captured_per_game": 59.479166666666664, "stones_lost_per_game": 11.4375, "passes_per_game": 3.25, "games_with_a_pass_before_move_60": 0, "end_two_passes": 87, "end_repetition": 8, "end_move_limit": 1}, "weak": {"games": 32, "score": 0.875, "wins": 28, "losses": 4, "score_as_black": 0.9375, "score_as_white": 0.8125, "mean_margin": 71.6875, "mean_length": 116.1875, "stones_captured_per_game": 76.59375, "stones_lost_per_game": 2.28125, "passes_per_game": 2.84375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 26, "end_repetition": 5, "end_move_limit": 1}}, "files": "/work/pass_21/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 21, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.75, "wins": 72, "losses": 24, "score_as_black": 0.7083333134651184, "score_as_white": 0.7916666865348816, "mean_margin": 48.375, "mean_length": 117.54166666666667, "stones_captured_per_game": 59.479166666666664, "stones_lost_per_game": 11.4375, "passes_per_game": 3.25, "games_with_a_pass_before_move_60": 0, "end_two_passes": 87, "end_repetition": 8, "end_move_limit": 1}, "weak": {"games": 32, "score": 0.875, "wins": 28, "losses": 4, "score_as_black": 0.9375, "score_as_white": 0.8125, "mean_margin": 71.6875, "mean_length": 116.1875, "stones_captured_per_game": 76.59375, "stones_lost_per_game": 2.28125, "passes_per_game": 2.84375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 26, "end_repetition": 5, "end_move_limit": 1}}, "files": "/work/pass_21/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 4.

python

import json,shutilshutil.copyfile("/work/design.py","/work/design_v21.py")s=open("/work/design_v19.py").read()idx=s.index('OB[81] = P["pass_bias"]')extra=r'''# Two witnessed solid enemy eyes warn against treating a large living# string as capturable merely because its liberties are far apart.eye_points=[r*9+c for r in range(0,9,2) for c in range(0,9,2)]for j,p in enumerate(eye_points):    u=487+j    ring=adj[p]+diag[p]    B[4][u]=1-len(ring)    for q in ring:        W[4][unit(1,q),u]+=1    W[4][unit(0,p),u]-=8    W[4][unit(1,p),u]-=8
B[5][487]=1for v in range(487,512):    W[5][v,487]=-1for p in range(81):    # Positive capture evidence, disabled when two solid eyes are seen.    u=unit(0,p)    W[5][:,u]=0    B[5][u]=4    for q in adj[p]:        W[5][unit(3,q),u]+=1    for v in range(487,512):        W[5][v,u]-=4    OW[u,p]=P["capture"]    OW[487,p]-=4*P["capture"]    OB[p]-=P["capture"]
'''s=s[:idx]+extra+s[idx:]open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added a two-solid-enemy-eyes safeguard against false capture urgency.")

import json,shutilshutil.copyfile("/work/design.py","/work/design_v21.py")s=open("/work/design_v19.py").read()idx=s.index('OB[81] = P["pass_bias"]')extra=r'''# Two witnessed solid enemy eyes warn against treating a large living# string as capturable merely because its liberties are far apart.eye_points=[r*9+c for r in range(0,9,2) for c in range(0,9,2)]for j,p in enumerate(eye_points):    u=487+j    ring=adj[p]+diag[p]    B[4][u]=1-len(ring)    for q in ring:        W[4][unit(1,q),u]+=1    W[4][unit(0,p),u]-=8    W[4][unit(1,p),u]-=8
B[5][487]=1for v in range(487,512):    W[5][v,487]=-1for p in range(81):    # Positive capture evidence, disabled when two solid eyes are seen.    u=unit(0,p)    W[5][:,u]=0    B[5][u]=4    for q in adj[p]:        W[5][unit(3,q),u]+=1    for v in range(487,512):        W[5][v,u]-=4    OW[u,p]=P["capture"]    OW[487,p]-=4*P["capture"]    OB[p]-=P["capture"]
'''s=s[:idx]+extra+s[idx:]open("/work/design.py","w").write(s)overrides=json.load(open("/work/params_v9.json"))exec(s)print("Added a two-solid-enemy-eyes safeguard against false capture urgency.")

OUTPUT of Python run 29 of at most 120:

Games pass 22 of 25 done.

{"ok": true, "pass": 22, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.6354166865348816, "wins": 61, "losses": 35, "score_as_black": 0.625, "score_as_white": 0.6458333134651184, "mean_margin": 44.03125, "mean_length": 110.69791666666667, "stones_captured_per_game": 55.354166666666664, "stones_lost_per_game": 11.479166666666666, "passes_per_game": 1.4791666666666667, "games_with_a_pass_before_move_60": 0, "end_two_passes": 79, "end_repetition": 16, "end_move_limit": 1}, "weak": {"games": 32, "score": 0.96875, "wins": 31, "losses": 1, "score_as_black": 0.9375, "score_as_white": 1.0, "mean_margin": 71.875, "mean_length": 117.5625, "stones_captured_per_game": 73.375, "stones_lost_per_game": 2.34375, "passes_per_game": 2.6875, "games_with_a_pass_before_move_60": 0, "end_two_passes": 32, "end_repetition": 0, "end_move_limit": 0}}, "files": "/work/pass_22/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 22, "kind": "games", "per_opponent": {"strong": {"games": 96, "score": 0.6354166865348816, "wins": 61, "losses": 35, "score_as_black": 0.625, "score_as_white": 0.6458333134651184, "mean_margin": 44.03125, "mean_length": 110.69791666666667, "stones_captured_per_game": 55.354166666666664, "stones_lost_per_game": 11.479166666666666, "passes_per_game": 1.4791666666666667, "games_with_a_pass_before_move_60": 0, "end_two_passes": 79, "end_repetition": 16, "end_move_limit": 1}, "weak": {"games": 32, "score": 0.96875, "wins": 31, "losses": 1, "score_as_black": 0.9375, "score_as_white": 1.0, "mean_margin": 71.875, "mean_length": 117.5625, "stones_captured_per_game": 73.375, "stones_lost_per_game": 2.34375, "passes_per_game": 2.6875, "games_with_a_pass_before_move_60": 0, "end_two_passes": 32, "end_repetition": 0, "end_move_limit": 0}}, "files": "/work/pass_22/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 3.

python

import json, shutil, numpy as npshutil.copyfile("/work/design.py","/work/design_v22.py")final_source=open("/work/design_v19.py").read()open("/work/design.py","w").write(final_source)open("/work/final_design.py","w").write(final_source)overrides=json.load(open("/work/params_v9.json"))open("/work/final_parameters.json","w").write(json.dumps(overrides,indent=2))exec(final_source)
# Audit tensors only; no policy evaluation is performed here.total=0for name in saved():    a=load(name)    assert np.isfinite(a).all(), name    total+=a.size    print(name, a.shape, "nonzero", np.count_nonzero(a),          "max_abs", float(np.max(np.abs(a))))assert total==1480274print("Final network restored and saved; all 1,480,274 parameters are finite.")

import json, shutil, numpy as npshutil.copyfile("/work/design.py","/work/design_v22.py")final_source=open("/work/design_v19.py").read()open("/work/design.py","w").write(final_source)open("/work/final_design.py","w").write(final_source)overrides=json.load(open("/work/params_v9.json"))open("/work/final_parameters.json","w").write(json.dumps(overrides,indent=2))exec(final_source)
# Audit tensors only; no policy evaluation is performed here.total=0for name in saved():    a=load(name)    assert np.isfinite(a).all(), name    total+=a.size    print(name, a.shape, "nonzero", np.count_nonzero(a),          "max_abs", float(np.max(np.abs(a))))assert total==1480274print("Final network restored and saved; all 1,480,274 parameters are finite.")

OUTPUT of Python run 30 of at most 120:

L1.W (243, 512) nonzero 3368 max_abs 16.0L1.b (512,) nonzero 349 max_abs 61.0L2.W (512, 512) nonzero 4203 max_abs 16.0L2.b (512,) nonzero 186 max_abs 61.0L3.W (512, 512) nonzero 3896 max_abs 16.0L3.b (512,) nonzero 105 max_abs 4.0L4.W (512, 512) nonzero 3234 max_abs 4096.0L4.b (512,) nonzero 186 max_abs 20.0L5.W (512, 512) nonzero 1895 max_abs 4096.0L5.b (512,) nonzero 0 max_abs 0.0L6.W (512, 512) nonzero 5548 max_abs 2.0L6.b (512,) nonzero 405 max_abs 3.5out.W (512, 82) nonzero 487 max_abs 25.600000381469727out.b (82,) nonzero 82 max_abs 14.080000877380371Final network restored and saved; all 1,480,274 parameters are finite.

L1.W (243, 512) nonzero 3368 max_abs 16.0L1.b (512,) nonzero 349 max_abs 61.0L2.W (512, 512) nonzero 4203 max_abs 16.0L2.b (512,) nonzero 186 max_abs 61.0L3.W (512, 512) nonzero 3896 max_abs 16.0L3.b (512,) nonzero 105 max_abs 4.0L4.W (512, 512) nonzero 3234 max_abs 4096.0L4.b (512,) nonzero 186 max_abs 20.0L5.W (512, 512) nonzero 1895 max_abs 4096.0L5.b (512,) nonzero 0 max_abs 0.0L6.W (512, 512) nonzero 5548 max_abs 2.0L6.b (512,) nonzero 405 max_abs 3.5out.W (512, 82) nonzero 487 max_abs 25.600000381469727out.b (82,) nonzero 82 max_abs 14.080000877380371Final network restored and saved; all 1,480,274 parameters are finite.

positions

{"positions":[{"board":[".........",".........",".........",".........",".........",".........",".........",".........","........."],"to_move":"black"},{"board":[".........",".........",".........","....X....","...XO....","....X....",".........",".........","........."],"to_move":"black"},{"board":[".........",".........",".........","....O....","...OX....","....O....",".........",".........","........."],"to_move":"black"},{"board":[".........",".........",".........","...XX....","..XOO....","...XX....",".........",".........","........."],"to_move":"black"},{"board":[".........",".........",".........","....X....","...X.X...","....X....",".........",".........","........."],"to_move":"black"},{"board":["XXXXXXXXX","XXXXXXXXX","XX.XXXXXX","XXXXXXXXX","XXXXXXXXX","XXXXXXXXX","XXXXXX.XX","XXXXXXXXX","XXXXXXXXX"],"to_move":"black"}],"activations":true}

{"positions":[{"board":[".........",".........",".........",".........",".........",".........",".........",".........","........."],"to_move":"black"},{"board":[".........",".........",".........","....X....","...XO....","....X....",".........",".........","........."],"to_move":"black"},{"board":[".........",".........",".........","....O....","...OX....","....O....",".........",".........","........."],"to_move":"black"},{"board":[".........",".........",".........","...XX....","..XOO....","...XX....",".........",".........","........."],"to_move":"black"},{"board":[".........",".........",".........","....X....","...X.X...","....X....",".........",".........","........."],"to_move":"black"},{"board":["XXXXXXXXX","XXXXXXXXX","XX.XXXXXX","XXXXXXXXX","XXXXXXXXX","XXXXXXXXX","XXXXXX.XX","XXXXXXXXX","XXXXXXXXX"],"to_move":"black"}],"activations":true}

Positions pass 23 of 25 done.

{"ok": true, "pass": 23, "kind": "positions", "positions": 6, "top3_per_position": [[["(4,4)", 1.0], ["(2,6)", 0.0], ["(6,6)", 0.0]], [["(4,5)", 1.0], ["(3,3)", 0.0], ["(5,3)", 0.0]], [["(4,5)", 0.9996], ["(3,3)", 0.0001], ["(5,3)", 0.0001]], [["(4,5)", 1.0], ["(3,2)", 0.0], ["(5,2)", 0.0]], [["(3,3)", 0.2062], ["(5,5)", 0.2062], ["(5,3)", 0.2062]], [["pass", 1.0], ["(2,2)", 0.0], ["(6,6)", 0.0]]], "files": "/work/pass_23/", "readme": "inputs.npy (P, 243) the network inputs; legal.npy (P, 82) the legal moves (for 'board' positions: the empty points and pass; suicide and ko are not checked); logits.npy, probs.npy (P, 82); act_L1.npy .. act_L6.npy (P, 512) the hidden activations after ReLU."}

{"ok": true, "pass": 23, "kind": "positions", "positions": 6, "top3_per_position": [[["(4,4)", 1.0], ["(2,6)", 0.0], ["(6,6)", 0.0]], [["(4,5)", 1.0], ["(3,3)", 0.0], ["(5,3)", 0.0]], [["(4,5)", 0.9996], ["(3,3)", 0.0001], ["(5,3)", 0.0001]], [["(4,5)", 1.0], ["(3,2)", 0.0], ["(5,2)", 0.0]], [["(3,3)", 0.2062], ["(5,5)", 0.2062], ["(5,3)", 0.2062]], [["pass", 1.0], ["(2,2)", 0.0], ["(6,6)", 0.0]]], "files": "/work/pass_23/", "readme": "inputs.npy (P, 243) the network inputs; legal.npy (P, 82) the legal moves (for 'board' positions: the empty points and pass; suicide and ko are not checked); logits.npy, probs.npy (P, 82); act_L1.npy .. act_L6.npy (P, 512) the hidden activations after ReLU."}

Passes left: 2.

Games pass 24 of 25 done.

{"ok": true, "pass": 24, "kind": "games", "per_opponent": {"strong": {"games": 128, "score": 0.734375, "wins": 94, "losses": 34, "score_as_black": 0.703125, "score_as_white": 0.765625, "mean_margin": 52.6171875, "mean_length": 116.671875, "stones_captured_per_game": 62.7109375, "stones_lost_per_game": 10.3359375, "passes_per_game": 1.671875, "games_with_a_pass_before_move_60": 0, "end_two_passes": 114, "end_repetition": 13, "end_move_limit": 1}}, "files": "/work/pass_24/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 24, "kind": "games", "per_opponent": {"strong": {"games": 128, "score": 0.734375, "wins": 94, "losses": 34, "score_as_black": 0.703125, "score_as_white": 0.765625, "mean_margin": 52.6171875, "mean_length": 116.671875, "stones_captured_per_game": 62.7109375, "stones_lost_per_game": 10.3359375, "passes_per_game": 1.671875, "games_with_a_pass_before_move_60": 0, "end_two_passes": 114, "end_repetition": 13, "end_move_limit": 1}}, "files": "/work/pass_24/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 1.

Games pass 25 of 25 done.

{"ok": true, "pass": 25, "kind": "games", "per_opponent": {"random": {"games": 64, "score": 0.984375, "wins": 63, "losses": 1, "score_as_black": 1.0, "score_as_white": 0.96875, "mean_margin": 77.84375, "mean_length": 117.21875, "stones_captured_per_game": 80.84375, "stones_lost_per_game": 0.5, "passes_per_game": 1.109375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 63, "end_repetition": 1, "end_move_limit": 0}, "weak": {"games": 64, "score": 0.921875, "wins": 59, "losses": 5, "score_as_black": 0.875, "score_as_white": 0.96875, "mean_margin": 71.796875, "mean_length": 118.8125, "stones_captured_per_game": 72.8125, "stones_lost_per_game": 4.265625, "passes_per_game": 2.125, "games_with_a_pass_before_move_60": 0, "end_two_passes": 60, "end_repetition": 4, "end_move_limit": 0}}, "files": "/work/pass_25/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

{"ok": true, "pass": 25, "kind": "games", "per_opponent": {"random": {"games": 64, "score": 0.984375, "wins": 63, "losses": 1, "score_as_black": 1.0, "score_as_white": 0.96875, "mean_margin": 77.84375, "mean_length": 117.21875, "stones_captured_per_game": 80.84375, "stones_lost_per_game": 0.5, "passes_per_game": 1.109375, "games_with_a_pass_before_move_60": 0, "end_two_passes": 63, "end_repetition": 1, "end_move_limit": 0}, "weak": {"games": 64, "score": 0.921875, "wins": 59, "losses": 5, "score_as_black": 0.875, "score_as_white": 0.96875, "mean_margin": 71.796875, "mean_length": 118.8125, "stones_captured_per_game": 72.8125, "stones_lost_per_game": 4.265625, "passes_per_game": 2.125, "games_with_a_pass_before_move_60": 0, "end_two_passes": 60, "end_repetition": 4, "end_move_limit": 0}}, "files": "/work/pass_25/", "readme": "games.json: one entry per game (opponent, net_color, result +1/-1, margin = net's area score minus the opponent's incl. komi, end, stones captured / lost, moves: ply, by, move (row*9+col, 81 = pass), and for the network's moves its top-5 [move, probability]). boards.npy: (games, 162, 81) int8, the board AFTER each ply (1 Black, -1 White, 0 empty; index row*9+col); valid.npy: (games, 162) which plies were played."}

Passes left: 0.

All weights are saved. The network is entirely hand-designed—no training or search.

Its ReLU layers implement:

Local liberty counting and group-safety propagation, strengthened by connected 2×2-block certificates.
Capture and rescue priorities.
Strong self-atari and eye-filling penalties.
Connection, blocking, contact-fighting, and crowding preferences.
A center opening and conservative passing.

Final validation:

Strong: 94/128 wins — 73.4%
Weak: 59/64 — 92.2%
Random: 63/64 — 98.4%

I expect its tactical discipline and connected fighting shapes to generalize reasonably well. Long-range life-and-death and repetition cycles remain its main weaknesses.