File size: 38,828 Bytes
22d295b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 |
#!/usr/bin/env python3
"""
GEOMETRIC CODEBOOK - Signal-to-Language Decoder
Ghost in the Machine Labs
The missing piece: converts fused substrate geometric outputs
into actionable text (analysis) and Python code (solve functions).
Architecture:
Grid → GeometricEncoder → substrate signal (numpy)
Substrate processes → output signal (numpy)
Output signal → GeometricDecoder → {analysis text, hypothesis, Python code}
The codebook works by detecting geometric PRIMITIVES in the substrate
output — the torsions, symmetries, and relational patterns that the
sensor panels identified — and mapping them to transformation OPERATIONS
that can be expressed as code.
This is fabrication, not training. Each primitive→operation mapping is
a direct geometric relationship, not a learned weight.
"""
import numpy as np
from typing import List, Dict, Tuple, Optional, Any
from dataclasses import dataclass, field
from enum import Enum
# =============================================================================
# GEOMETRIC PRIMITIVES — the vocabulary of transformations
# =============================================================================
class GeoPrimitive(Enum):
"""
Primitives detected by substrate sensor panels.
Each maps to one or more ARC transformation operations.
"""
# Spatial
TILE_REPEAT = "tile_repeat" # Pattern repeats in space
MIRROR_H = "mirror_horizontal" # Horizontal reflection
MIRROR_V = "mirror_vertical" # Vertical reflection
MIRROR_DIAG = "mirror_diagonal" # Diagonal reflection
ROTATE_90 = "rotate_90"
ROTATE_180 = "rotate_180"
ROTATE_270 = "rotate_270"
TRANSLATE = "translate" # Shift pattern
SCALE_UP = "scale_up" # Enlarge grid/pattern
SCALE_DOWN = "scale_down" # Shrink grid/pattern
# Chromatic (color operations)
COLOR_MAP = "color_map" # Recolor by mapping
COLOR_FILL = "color_fill" # Flood fill region
COLOR_SWAP = "color_swap" # Swap two colors
COLOR_COUNT = "color_count" # Count-based operation
MAJORITY_COLOR = "majority_color" # Most frequent color
BOUNDARY_COLOR = "boundary_color" # Color at edges
# Structural
EXTRACT_SHAPE = "extract_shape" # Pull out a shape
MASK_OVERLAY = "mask_overlay" # Apply mask/overlay
CROP = "crop" # Trim to content
PAD = "pad" # Add border
PARTITION = "partition" # Split into regions
GRAVITY = "gravity" # Objects fall/slide
# Relational
SORT_BY_SIZE = "sort_by_size" # Order by region size
ALIGN = "align" # Align objects
CONNECT = "connect" # Draw lines between
ENCLOSE = "enclose" # Draw boundary around
# Pattern
DENOISING = "denoising" # Remove noise
COMPLETE_PATTERN = "complete_pattern" # Fill in missing
BOOLEAN_OP = "boolean_op" # AND/OR/XOR grids
CONDITIONAL = "conditional" # If-then color rules
@dataclass
class DetectedPrimitive:
"""A primitive detected in the substrate output with confidence."""
primitive: GeoPrimitive
confidence: float
params: Dict[str, Any] = field(default_factory=dict)
# Geometric evidence from substrate
sensor_source: str = "" # Which sensor type detected it
energy: float = 0.0 # Signal energy at detection
harmonic_alignment: float = 0.0 # How aligned with field
# =============================================================================
# GRID ENCODER — ARC grids to substrate signals
# =============================================================================
class GridEncoder:
"""
Encode ARC grids as geometric signals for the substrate.
Rather than flattening to bytes, we encode the GEOMETRIC PROPERTIES
of the grid that the sensor panels are designed to detect:
- Spatial frequency (tiling/repetition)
- Color distribution (chromatic spectrum)
- Shape boundaries (structural edges)
- Symmetry axes
- Relational positions
"""
SIGNAL_SIZE = 1024
@staticmethod
def encode_grid(grid: List[List[int]]) -> np.ndarray:
"""Encode a single grid into geometric signal."""
g = np.array(grid, dtype=np.float32)
h, w = g.shape
signal = np.zeros(GridEncoder.SIGNAL_SIZE, dtype=np.float32)
idx = 0
# === Band 1: Shape signature (0-127) ===
# Dimensions encoded as ratios
signal[idx] = h / 30.0; idx += 1
signal[idx] = w / 30.0; idx += 1
signal[idx] = h * w / 900.0; idx += 1
signal[idx] = h / w if w > 0 else 1.0; idx += 1
# Flattened grid (normalized to 0-1)
flat = g.flatten() / 9.0
n = min(len(flat), 124)
signal[idx:idx+n] = flat[:n]
idx = 128
# === Band 2: Color spectrum (128-255) ===
# Color histogram (10 colors, 0-9)
for c in range(10):
count = np.sum(g == c)
signal[idx + c] = count / (h * w)
idx += 10
# Color adjacency matrix (which colors touch which)
for r in range(h):
for c in range(w):
val = int(g[r, c])
# Right neighbor
if c + 1 < w:
n_val = int(g[r, c + 1])
if val != n_val:
pair_idx = 128 + 10 + val * 10 + n_val
if pair_idx < 256:
signal[pair_idx] += 1.0 / (h * w)
# Down neighbor
if r + 1 < h:
n_val = int(g[r + 1, c])
if val != n_val:
pair_idx = 128 + 10 + val * 10 + n_val
if pair_idx < 256:
signal[pair_idx] += 1.0 / (h * w)
idx = 256
# === Band 3: Symmetry signatures (256-383) ===
# Horizontal symmetry
if h > 1:
h_sym = np.mean(g == g[::-1, :])
signal[idx] = h_sym
idx += 1
# Vertical symmetry
if w > 1:
v_sym = np.mean(g == g[:, ::-1])
signal[idx] = v_sym
idx += 1
# Diagonal symmetry (if square)
if h == w:
d_sym = np.mean(g == g.T)
signal[idx] = d_sym
idx += 1
# 90° rotational symmetry (if square)
if h == w:
r90 = np.rot90(g)
signal[idx] = np.mean(g == r90)
idx += 1
idx = 384
# === Band 4: Spatial frequency (384-511) ===
# Row-wise and column-wise repetition patterns
for r in range(min(h, 30)):
row = g[r, :]
# Check for period-N repetition
for period in range(1, min(w, 8)):
if w % period == 0:
tiles = row.reshape(-1, period)
if len(tiles) > 1 and np.all(tiles == tiles[0]):
signal[384 + r * 4 + min(period-1, 3)] = 1.0
break
idx = 512
# === Band 5: Boundary/edge features (512-639) ===
# Edge detection (Sobel-like)
if h > 2 and w > 2:
for r in range(1, min(h-1, 16)):
for c in range(1, min(w-1, 8)):
# Gradient magnitude
gx = float(g[r, c+1]) - float(g[r, c-1])
gy = float(g[r+1, c]) - float(g[r-1, c])
mag = np.sqrt(gx**2 + gy**2)
eidx = 512 + r * 8 + c
if eidx < 640:
signal[eidx] = mag / 12.73 # max gradient = 9*sqrt(2)
idx = 640
# === Band 6: Object detection (640-767) ===
# Connected component count per color
visited = np.zeros_like(g, dtype=bool)
obj_count = 0
for r in range(h):
for c in range(w):
if not visited[r, c]:
color = g[r, c]
if color != 0: # Skip background
# BFS
stack = [(r, c)]
size = 0
while stack:
cr, cc = stack.pop()
if 0 <= cr < h and 0 <= cc < w and not visited[cr, cc] and g[cr, cc] == color:
visited[cr, cc] = True
size += 1
stack.extend([(cr+1,cc),(cr-1,cc),(cr,cc+1),(cr,cc-1)])
if size > 0 and 640 + obj_count < 768:
signal[640 + obj_count] = size / (h * w)
obj_count += 1
signal[767] = obj_count / 30.0 # total object count
idx = 768
# === Band 7: Transformation hints (768-895) ===
# Reserved for encoding input→output relationships
# (filled by encode_pair)
idx = 896
# === Band 8: Raw hash (896-1023) ===
# Deterministic hash for exact matching
raw = g.tobytes()
for i in range(min(128, len(raw))):
signal[896 + i] = (raw[i] - 128.0) / 128.0
return signal
@staticmethod
def encode_pair(input_grid: List[List[int]],
output_grid: List[List[int]]) -> np.ndarray:
"""
Encode an input→output pair, capturing the TRANSFORMATION.
Band 7 (768-895) encodes the relationship:
- Dimension change ratios
- Color mapping
- Spatial operation signature
"""
sig = GridEncoder.encode_grid(input_grid)
ig = np.array(input_grid, dtype=np.float32)
og = np.array(output_grid, dtype=np.float32)
ih, iw = ig.shape
oh, ow = og.shape
idx = 768
# Dimension relationships
sig[idx] = oh / ih if ih > 0 else 1.0; idx += 1 # height ratio
sig[idx] = ow / iw if iw > 0 else 1.0; idx += 1 # width ratio
sig[idx] = (oh * ow) / (ih * iw) if ih * iw > 0 else 1.0; idx += 1 # area ratio
sig[idx] = 1.0 if oh == ih and ow == iw else 0.0; idx += 1 # same size?
# Color transformation
in_colors = set(ig.flatten().astype(int))
out_colors = set(og.flatten().astype(int))
sig[idx] = len(in_colors) / 10.0; idx += 1
sig[idx] = len(out_colors) / 10.0; idx += 1
sig[idx] = len(in_colors & out_colors) / max(len(in_colors | out_colors), 1); idx += 1
sig[idx] = 1.0 if in_colors == out_colors else 0.0; idx += 1
# If same size, compute cell-wise diff
if ih == oh and iw == ow:
diff = (ig != og).astype(np.float32)
sig[idx] = np.mean(diff); idx += 1 # fraction changed
sig[idx] = np.sum(diff) / max(ih * iw, 1); idx += 1 # changed count ratio
# Where did changes happen? Edge vs center
if ih > 2 and iw > 2:
edge_mask = np.zeros_like(diff)
edge_mask[0, :] = 1; edge_mask[-1, :] = 1
edge_mask[:, 0] = 1; edge_mask[:, -1] = 1
edge_changes = np.sum(diff * edge_mask)
center_changes = np.sum(diff * (1 - edge_mask))
total_changes = edge_changes + center_changes
sig[idx] = edge_changes / max(total_changes, 1); idx += 1
sig[idx] = center_changes / max(total_changes, 1); idx += 1
else:
idx += 4
# Tiling check: does output = tiled input?
if oh > ih and ow > iw and oh % ih == 0 and ow % iw == 0:
tile_h = oh // ih
tile_w = ow // iw
tiled = np.tile(ig, (tile_h, tile_w))
sig[idx] = np.mean(tiled == og); idx += 1 # simple tile match
sig[idx] = tile_h / 10.0; idx += 1
sig[idx] = tile_w / 10.0; idx += 1
# Check alternating tile (flip every other)
alt_tiled = np.zeros_like(og)
for tr in range(tile_h):
for tc in range(tile_w):
block = ig.copy()
if tr % 2 == 1:
block = block[::-1, :]
if tc % 2 == 1:
block = block[:, ::-1]
alt_tiled[tr*ih:(tr+1)*ih, tc*iw:(tc+1)*iw] = block
sig[idx] = np.mean(alt_tiled == og); idx += 1 # alternating tile match
else:
idx += 4
# Rotation check (if square)
if ih == iw:
r90 = np.rot90(ig)
r180 = np.rot90(ig, 2)
r270 = np.rot90(ig, 3)
if oh == ih and ow == iw:
sig[idx] = np.mean(r90 == og); idx += 1
sig[idx] = np.mean(r180 == og); idx += 1
sig[idx] = np.mean(r270 == og); idx += 1
else:
idx += 3
else:
idx += 3
# Mirror check
if oh == ih and ow == iw:
sig[idx] = np.mean(ig[::-1, :] == og); idx += 1 # flip H
sig[idx] = np.mean(ig[:, ::-1] == og); idx += 1 # flip V
return sig
@staticmethod
def encode_task(task: Dict) -> np.ndarray:
"""
Encode full ARC task (all training pairs) into composite signal.
Averages pair signals to find the CONSISTENT transformation
pattern across all examples.
"""
train = task.get('train', [])
if not train:
return np.zeros(GridEncoder.SIGNAL_SIZE, dtype=np.float32)
signals = []
for pair in train:
sig = GridEncoder.encode_pair(pair['input'], pair['output'])
signals.append(sig)
# Consensus: average across pairs
# High-confidence features will be consistent, noise will cancel
composite = np.mean(signals, axis=0).astype(np.float32)
# Variance across pairs (low = consistent = confident)
if len(signals) > 1:
variance = np.var(signals, axis=0)
# Boost consistent features, dampen noisy ones
consistency = 1.0 / (1.0 + variance * 10)
composite *= consistency
return composite
# =============================================================================
# PRIMITIVE DETECTOR — reads substrate output to find operations
# =============================================================================
class PrimitiveDetector:
"""
Detect geometric primitives from substrate output signal.
The substrate's sensor panels have already done the hard work —
detecting spatial frequencies, symmetries, boundaries, etc.
The detector reads those activations and maps them to named
transformation primitives.
"""
# Thresholds for detection
CONFIDENCE_THRESHOLD = 0.3
@staticmethod
def detect(input_signal: np.ndarray,
output_signal: np.ndarray,
task: Dict) -> List[DetectedPrimitive]:
"""
Detect all primitives from substrate processing results.
Uses both the encoded signal AND the raw task data to
cross-validate detections.
"""
primitives = []
train = task.get('train', [])
if not train:
return primitives
# Analyze all training pairs for consensus
pair_primitives = []
for pair in train:
pp = PrimitiveDetector._detect_pair(pair['input'], pair['output'])
pair_primitives.append(set(p.primitive for p in pp))
primitives.extend(pp)
# Consensus: only keep primitives detected in ALL pairs
if pair_primitives:
consensus = pair_primitives[0]
for ps in pair_primitives[1:]:
consensus &= ps
if consensus:
# Filter to consensus + boost confidence
consensus_prims = []
for p in primitives:
if p.primitive in consensus:
p.confidence = min(1.0, p.confidence * 1.5) # boost
consensus_prims.append(p)
# Deduplicate: keep highest confidence per primitive type
best = {}
for p in consensus_prims:
if p.primitive not in best or p.confidence > best[p.primitive].confidence:
best[p.primitive] = p
primitives = list(best.values())
else:
# No perfect consensus — keep all and deduplicate by confidence
best = {}
for p in primitives:
if p.primitive not in best or p.confidence > best[p.primitive].confidence:
best[p.primitive] = p
primitives = list(best.values())
# Sort by confidence
primitives.sort(key=lambda p: p.confidence, reverse=True)
return primitives
@staticmethod
def _detect_pair(input_grid: List[List[int]],
output_grid: List[List[int]]) -> List[DetectedPrimitive]:
"""Detect primitives for a single input→output pair."""
prims = []
ig = np.array(input_grid, dtype=np.float32)
og = np.array(output_grid, dtype=np.float32)
ih, iw = ig.shape
oh, ow = og.shape
# --- TILING ---
if oh > ih and ow > iw and oh % ih == 0 and ow % iw == 0:
tile_h, tile_w = oh // ih, ow // iw
# Simple tile
tiled = np.tile(ig, (tile_h, tile_w))
match = np.mean(tiled == og)
if match > 0.8:
prims.append(DetectedPrimitive(
GeoPrimitive.TILE_REPEAT, match,
{'tile_h': tile_h, 'tile_w': tile_w, 'alternating': False},
'spatial'))
# Alternating tile: tile columns, alternate block-rows
# between original and column-reversed
alt = np.zeros_like(og)
for tr in range(tile_h):
block = ig.copy()
if tr % 2 == 1:
block = block[:, ::-1] # reverse columns on odd blocks
row_tile = np.tile(block, (1, tile_w))
alt[tr*ih:(tr+1)*ih, :] = row_tile
alt_match = np.mean(alt == og)
if alt_match > match and alt_match > 0.8:
prims.append(DetectedPrimitive(
GeoPrimitive.TILE_REPEAT, alt_match,
{'tile_h': tile_h, 'tile_w': tile_w, 'alternating': True},
'spatial'))
# --- SCALING ---
if oh > ih and ow > iw:
sh, sw = oh / ih, ow / iw
if sh == sw and sh == int(sh):
scale = int(sh)
scaled = np.repeat(np.repeat(ig, scale, axis=0), scale, axis=1)
if scaled.shape == og.shape:
match = np.mean(scaled == og)
if match > 0.8:
prims.append(DetectedPrimitive(
GeoPrimitive.SCALE_UP, match,
{'factor': scale}, 'spatial'))
if oh < ih and ow < iw and ih % oh == 0 and iw % ow == 0:
prims.append(DetectedPrimitive(
GeoPrimitive.SCALE_DOWN, 0.7,
{'factor_h': ih // oh, 'factor_w': iw // ow}, 'spatial'))
# --- ROTATION (same size, square) ---
if ih == iw and oh == ow and ih == oh:
for k, prim in [(1, GeoPrimitive.ROTATE_90),
(2, GeoPrimitive.ROTATE_180),
(3, GeoPrimitive.ROTATE_270)]:
rotated = np.rot90(ig, k)
match = np.mean(rotated == og)
if match > 0.9:
prims.append(DetectedPrimitive(prim, match, {}, 'symmetry'))
# --- MIRROR ---
if ih == oh and iw == ow:
# Horizontal flip
flipped_h = ig[::-1, :]
match_h = np.mean(flipped_h == og)
if match_h > 0.9:
prims.append(DetectedPrimitive(
GeoPrimitive.MIRROR_H, match_h, {}, 'symmetry'))
# Vertical flip
flipped_v = ig[:, ::-1]
match_v = np.mean(flipped_v == og)
if match_v > 0.9:
prims.append(DetectedPrimitive(
GeoPrimitive.MIRROR_V, match_v, {}, 'symmetry'))
# Transpose
if ih == iw:
transposed = ig.T
match_t = np.mean(transposed == og)
if match_t > 0.9:
prims.append(DetectedPrimitive(
GeoPrimitive.MIRROR_DIAG, match_t, {}, 'symmetry'))
# --- COLOR MAP ---
if ih == oh and iw == ow:
# Check if there's a consistent color→color mapping
mapping = {}
consistent = True
for r in range(ih):
for c in range(iw):
ic = int(ig[r, c])
oc = int(og[r, c])
if ic in mapping:
if mapping[ic] != oc:
consistent = False
break
else:
mapping[ic] = oc
if not consistent:
break
if consistent and mapping:
is_identity = all(k == v for k, v in mapping.items())
if not is_identity:
prims.append(DetectedPrimitive(
GeoPrimitive.COLOR_MAP, 1.0,
{'mapping': mapping}, 'chromatic'))
# --- COLOR SWAP ---
if ih == oh and iw == ow:
diff_positions = ig != og
changed_in = set(ig[diff_positions].astype(int).tolist())
changed_out = set(og[diff_positions].astype(int).tolist())
if len(changed_in) == 2 and changed_in == changed_out:
colors = list(changed_in)
prims.append(DetectedPrimitive(
GeoPrimitive.COLOR_SWAP, 0.95,
{'color_a': colors[0], 'color_b': colors[1]}, 'chromatic'))
# --- CROP ---
if oh < ih or ow < iw:
# Check if output is a sub-region of input
for r in range(ih - oh + 1):
for c in range(iw - ow + 1):
region = ig[r:r+oh, c:c+ow]
if np.array_equal(region, og):
prims.append(DetectedPrimitive(
GeoPrimitive.CROP, 1.0,
{'top': r, 'left': c}, 'structural'))
break
else:
continue
break
# --- GRAVITY ---
if ih == oh and iw == ow:
# Check if non-zero cells "fell" downward
for c in range(iw):
in_col = ig[:, c]
out_col = og[:, c]
in_vals = in_col[in_col != 0]
out_vals = out_col[out_col != 0]
if len(in_vals) > 0 and np.array_equal(sorted(in_vals), sorted(out_vals)):
# Check if output has all non-zero at bottom
out_nonzero = np.where(out_col != 0)[0]
if len(out_nonzero) > 0 and out_nonzero[-1] == ih - 1:
if np.all(np.diff(out_nonzero) == 1):
prims.append(DetectedPrimitive(
GeoPrimitive.GRAVITY, 0.8,
{'direction': 'down'}, 'structural'))
break
# --- BOOLEAN OP ---
# (Detects when output = some combination of input regions)
# This is complex — will expand in later versions
return prims
# =============================================================================
# CODE GENERATOR — primitives to Python solve()
# =============================================================================
class CodeGenerator:
"""
Generate Python solve() functions from detected primitives.
Each primitive has a direct code template. Compositions
chain templates together.
"""
TEMPLATES = {
GeoPrimitive.TILE_REPEAT: {
'simple': """def solve(input_grid):
grid = np.array(input_grid)
h, w = grid.shape
tile_h, tile_w = {tile_h}, {tile_w}
result = np.tile(grid, (tile_h, tile_w))
return result.tolist()""",
'alternating': """def solve(input_grid):
grid = np.array(input_grid)
h, w = grid.shape
tile_h, tile_w = {tile_h}, {tile_w}
result = np.zeros((h * tile_h, w * tile_w), dtype=int)
for tr in range(tile_h):
block = grid.copy()
if tr % 2 == 1:
block = block[:, ::-1]
row_tile = np.tile(block, (1, tile_w))
result[tr*h:(tr+1)*h, :] = row_tile
return result.tolist()""",
},
GeoPrimitive.SCALE_UP: """def solve(input_grid):
grid = np.array(input_grid)
factor = {factor}
result = np.repeat(np.repeat(grid, factor, axis=0), factor, axis=1)
return result.tolist()""",
GeoPrimitive.ROTATE_90: """def solve(input_grid):
grid = np.array(input_grid)
result = np.rot90(grid, 1)
return result.tolist()""",
GeoPrimitive.ROTATE_180: """def solve(input_grid):
grid = np.array(input_grid)
result = np.rot90(grid, 2)
return result.tolist()""",
GeoPrimitive.ROTATE_270: """def solve(input_grid):
grid = np.array(input_grid)
result = np.rot90(grid, 3)
return result.tolist()""",
GeoPrimitive.MIRROR_H: """def solve(input_grid):
grid = np.array(input_grid)
result = grid[::-1, :]
return result.tolist()""",
GeoPrimitive.MIRROR_V: """def solve(input_grid):
grid = np.array(input_grid)
result = grid[:, ::-1]
return result.tolist()""",
GeoPrimitive.MIRROR_DIAG: """def solve(input_grid):
grid = np.array(input_grid)
result = grid.T
return result.tolist()""",
GeoPrimitive.COLOR_MAP: """def solve(input_grid):
grid = np.array(input_grid)
result = grid.copy()
mapping = {mapping}
for old_c, new_c in mapping.items():
result[grid == old_c] = new_c
return result.tolist()""",
GeoPrimitive.COLOR_SWAP: """def solve(input_grid):
grid = np.array(input_grid)
result = grid.copy()
a, b = {color_a}, {color_b}
result[grid == a] = b
result[grid == b] = a
return result.tolist()""",
GeoPrimitive.CROP: """def solve(input_grid):
grid = np.array(input_grid)
# Find bounding box of non-zero content
rows = np.any(grid != 0, axis=1)
cols = np.any(grid != 0, axis=0)
if not rows.any():
return input_grid
rmin, rmax = np.where(rows)[0][[0, -1]]
cmin, cmax = np.where(cols)[0][[0, -1]]
result = grid[rmin:rmax+1, cmin:cmax+1]
return result.tolist()""",
GeoPrimitive.GRAVITY: """def solve(input_grid):
grid = np.array(input_grid)
h, w = grid.shape
result = np.zeros_like(grid)
for c in range(w):
col = grid[:, c]
nonzero = col[col != 0]
result[h-len(nonzero):h, c] = nonzero
return result.tolist()""",
}
@staticmethod
def generate(primitives: List[DetectedPrimitive]) -> Optional[str]:
"""Generate solve() function from detected primitives."""
if not primitives:
return None
# Take highest confidence primitive
best = primitives[0]
template = CodeGenerator.TEMPLATES.get(best.primitive)
if template is None:
return None
# Handle templates with variants
if isinstance(template, dict):
if best.params.get('alternating'):
template = template.get('alternating', template.get('simple'))
else:
template = template.get('simple')
# Fill parameters
try:
code = template.format(**best.params)
except (KeyError, IndexError):
code = template
# Ensure numpy import
if 'np.' in code and 'import numpy' not in code:
code = "import numpy as np\n\n" + code
return code
@staticmethod
def generate_analysis(primitives: List[DetectedPrimitive],
task: Dict) -> str:
"""Generate text analysis from detected primitives."""
if not primitives:
return "No transformation pattern detected with sufficient confidence."
train = task.get('train', [])
ig = np.array(train[0]['input']) if train else np.array([[]])
og = np.array(train[0]['output']) if train else np.array([[]])
lines = []
lines.append(f"Input: {ig.shape[0]}x{ig.shape[1]} → Output: {og.shape[0]}x{og.shape[1]}")
for p in primitives:
desc = f"Detected {p.primitive.value} (confidence: {p.confidence:.2f})"
if p.params:
params_str = ", ".join(f"{k}={v}" for k, v in p.params.items())
desc += f" [{params_str}]"
lines.append(desc)
# Primary transformation description
best = primitives[0]
descriptions = {
GeoPrimitive.TILE_REPEAT: "The output tiles the input pattern across a larger grid",
GeoPrimitive.SCALE_UP: "The output scales up each cell of the input",
GeoPrimitive.ROTATE_90: "The output is the input rotated 90° counterclockwise",
GeoPrimitive.ROTATE_180: "The output is the input rotated 180°",
GeoPrimitive.ROTATE_270: "The output is the input rotated 270° counterclockwise",
GeoPrimitive.MIRROR_H: "The output flips the input horizontally (upside down)",
GeoPrimitive.MIRROR_V: "The output flips the input vertically (left-right)",
GeoPrimitive.MIRROR_DIAG: "The output transposes the input (diagonal mirror)",
GeoPrimitive.COLOR_MAP: "Each color in the input maps to a specific color in the output",
GeoPrimitive.COLOR_SWAP: "Two specific colors are swapped",
GeoPrimitive.CROP: "The output extracts a sub-region from the input",
GeoPrimitive.GRAVITY: "Non-zero cells fall to the bottom of each column",
}
lines.append(f"\nRule: {descriptions.get(best.primitive, best.primitive.value)}")
if best.params:
lines.append(f"Parameters: {best.params}")
return "\n".join(lines)
@staticmethod
def generate_hypothesis(primitives: List[DetectedPrimitive]) -> str:
"""Generate algorithm hypothesis from detected primitives."""
if not primitives:
return "Insufficient geometric evidence to form hypothesis."
best = primitives[0]
hypotheses = {
GeoPrimitive.TILE_REPEAT: "Tile the input grid {tile_h}x{tile_w} times. "
"If alternating: flip vertically on odd rows, horizontally on odd columns.",
GeoPrimitive.SCALE_UP: "Scale each cell to a {factor}x{factor} block.",
GeoPrimitive.ROTATE_90: "Apply np.rot90(grid, 1).",
GeoPrimitive.ROTATE_180: "Apply np.rot90(grid, 2).",
GeoPrimitive.ROTATE_270: "Apply np.rot90(grid, 3).",
GeoPrimitive.MIRROR_H: "Flip grid vertically: grid[::-1, :]",
GeoPrimitive.MIRROR_V: "Flip grid horizontally: grid[:, ::-1]",
GeoPrimitive.MIRROR_DIAG: "Transpose: grid.T",
GeoPrimitive.COLOR_MAP: "Apply color mapping: {mapping}",
GeoPrimitive.COLOR_SWAP: "Swap colors {color_a} ↔ {color_b}",
GeoPrimitive.CROP: "Extract bounding box of non-zero content.",
GeoPrimitive.GRAVITY: "For each column, move non-zero cells to bottom.",
}
template = hypotheses.get(best.primitive, str(best.primitive.value))
try:
return template.format(**best.params)
except (KeyError, IndexError):
return template
# =============================================================================
# INTEGRATED DECODER — full pipeline
# =============================================================================
class GeometricDecoder:
"""
Complete signal→text/code decoder.
This replaces _signal_to_response in the fused service.
"""
def __init__(self):
self.encoder = GridEncoder()
self.detector = PrimitiveDetector()
self.generator = CodeGenerator()
def decode_for_analysis(self, task: Dict,
substrate_output: np.ndarray) -> str:
"""Decode substrate output to text analysis."""
input_signal = self.encoder.encode_task(task)
primitives = self.detector.detect(input_signal, substrate_output, task)
return self.generator.generate_analysis(primitives, task)
def decode_for_hypothesis(self, task: Dict,
substrate_output: np.ndarray) -> str:
"""Decode substrate output to algorithm hypothesis."""
input_signal = self.encoder.encode_task(task)
primitives = self.detector.detect(input_signal, substrate_output, task)
return self.generator.generate_hypothesis(primitives)
def decode_for_code(self, task: Dict,
substrate_output: np.ndarray) -> Optional[str]:
"""Decode substrate output to Python solve() function."""
input_signal = self.encoder.encode_task(task)
primitives = self.detector.detect(input_signal, substrate_output, task)
return self.generator.generate(primitives)
def solve_task(self, task: Dict) -> Optional[str]:
"""
Full pipeline: task → encode → detect → code.
Bypasses substrate for direct geometric solving.
This is the one-pass fabrication path.
"""
train = task.get('train', [])
if not train:
return None
# Detect primitives directly from task geometry
primitives = self.detector.detect(
np.zeros(1024), np.zeros(1024), task)
if not primitives:
return None
return self.generator.generate(primitives)
# =============================================================================
# STANDALONE TEST
# =============================================================================
def test_codebook():
"""Test the codebook against known ARC patterns."""
decoder = GeometricDecoder()
# Test 1: Tiling (task 00576224)
task_tile = {
'train': [
{'input': [[7, 9], [4, 3]],
'output': [[7,9,7,9,7,9],[4,3,4,3,4,3],
[9,7,9,7,9,7],[3,4,3,4,3,4],
[7,9,7,9,7,9],[4,3,4,3,4,3]]},
{'input': [[8, 6], [6, 4]],
'output': [[8,6,8,6,8,6],[6,4,6,4,6,4],
[6,8,6,8,6,8],[4,6,4,6,4,6],
[8,6,8,6,8,6],[6,4,6,4,6,4]]},
],
'test': [
{'input': [[3, 2], [7, 8]],
'output': [[3,2,3,2,3,2],[7,8,7,8,7,8],
[2,3,2,3,2,3],[8,7,8,7,8,7],
[3,2,3,2,3,2],[7,8,7,8,7,8]]},
]
}
print("=" * 70)
print(" GEOMETRIC CODEBOOK TEST")
print("=" * 70)
# Test analysis
print("\n--- Task: Alternating Tile (00576224) ---")
analysis = decoder.decode_for_analysis(task_tile, np.zeros(1024))
print(f"Analysis:\n{analysis}")
hypothesis = decoder.decode_for_hypothesis(task_tile, np.zeros(1024))
print(f"\nHypothesis: {hypothesis}")
code = decoder.decode_for_code(task_tile, np.zeros(1024))
print(f"\nGenerated code:\n{code}")
# Validate
if code:
print("\n--- Validation ---")
namespace = {'np': np}
exec(code, namespace)
solve = namespace['solve']
for i, test in enumerate(task_tile['test']):
result = solve(test['input'])
expected = test['output']
match = result == expected
print(f" Test {i+1}: {'PASS ✓' if match else 'FAIL ✗'}")
if not match:
print(f" Expected: {expected[:2]}...")
print(f" Got: {result[:2]}...")
# Test 2: Simple rotation
print("\n--- Task: 90° Rotation ---")
task_rot = {
'train': [
{'input': [[1, 2], [3, 4]],
'output': [[2, 4], [1, 3]]},
{'input': [[5, 6], [7, 8]],
'output': [[6, 8], [5, 7]]},
],
'test': [
{'input': [[9, 1], [2, 3]],
'output': [[1, 3], [9, 2]]},
]
}
code = decoder.solve_task(task_rot)
if code:
print(f"Code: {code.strip().split(chr(10))[-1]}")
namespace = {'np': np}
exec(code, namespace)
result = namespace['solve'](task_rot['test'][0]['input'])
match = result == task_rot['test'][0]['output']
print(f" Test: {'PASS ✓' if match else 'FAIL ✗'}")
# Test 3: Color map
print("\n--- Task: Color Mapping ---")
task_color = {
'train': [
{'input': [[1, 2, 3], [1, 2, 3]],
'output': [[4, 5, 6], [4, 5, 6]]},
{'input': [[3, 1, 2], [2, 3, 1]],
'output': [[6, 4, 5], [5, 6, 4]]},
],
'test': [
{'input': [[2, 1, 3], [3, 2, 1]],
'output': [[5, 4, 6], [6, 5, 4]]},
]
}
code = decoder.solve_task(task_color)
if code:
print(f"Detected mapping")
namespace = {'np': np}
exec(code, namespace)
result = namespace['solve'](task_color['test'][0]['input'])
match = result == task_color['test'][0]['output']
print(f" Test: {'PASS ✓' if match else 'FAIL ✗'}")
print("\n" + "=" * 70)
if __name__ == "__main__":
test_codebook()
|