96e5f46866
Protects structure-mapping invariant. S1-S4 labels isolated from graph embeddings. Procrustes tests verify geometric separability and invariance. Verdict: GO.
141 lines
4 KiB
Python
141 lines
4 KiB
Python
import json
|
|
import numpy as np
|
|
from packs.compiler import load_mounted_packs
|
|
from core.physics.quantity_kernel import embed_quantity
|
|
from core.physics.dynamic_manifold import conformal_procrustes
|
|
import sys
|
|
|
|
def embed_graph(g_dict, vocab):
|
|
vectors = []
|
|
|
|
# 1. Initial State
|
|
for init in g_dict.get("initial_state", []):
|
|
qty = init["quantity"]["value"]
|
|
v_qty = embed_quantity(float(qty))
|
|
|
|
try:
|
|
v_kind = vocab.get_versor("have")
|
|
except Exception:
|
|
v_kind = np.zeros(32, dtype=np.float64)
|
|
v_kind[0] = 1.0
|
|
|
|
vectors.append(v_kind)
|
|
vectors.append(v_qty)
|
|
|
|
# 2. Operations
|
|
for op in g_dict.get("operations", []):
|
|
kind = op["kind"]
|
|
word_map = {
|
|
"compare_multiplicative": "multiply",
|
|
"compare_additive": "add",
|
|
"transfer": "give",
|
|
"add": "add",
|
|
"subtract": "subtract"
|
|
}
|
|
word = word_map.get(kind, kind)
|
|
try:
|
|
v_kind = vocab.get_versor(word)
|
|
except Exception:
|
|
v_kind = np.zeros(32, dtype=np.float64)
|
|
v_kind[0] = 1.0
|
|
v_kind[1] = 1.0
|
|
|
|
operand = op.get("operand", {})
|
|
val = 1.0
|
|
if "value" in operand:
|
|
val = operand["value"]
|
|
elif "factor" in operand:
|
|
val = operand["factor"]
|
|
|
|
v_qty = embed_quantity(float(val))
|
|
|
|
vectors.append(v_kind)
|
|
vectors.append(v_qty)
|
|
|
|
MAX_VECS = 10
|
|
padded = vectors[:MAX_VECS]
|
|
padding_point = embed_quantity(0.0)
|
|
while len(padded) < MAX_VECS:
|
|
padded.append(padding_point)
|
|
|
|
return padded
|
|
|
|
def main():
|
|
print("Loading vocab...")
|
|
# Load base math vocab
|
|
vocab = load_mounted_packs(["en_mathematics_logic_v1", "en_core_math_v1", "en_core_relations_v1", "en_core_relations_v2", "en_core_relations_v3"])
|
|
|
|
graphs_path = "sme_graphs.jsonl"
|
|
labels_path = "labels.jsonl"
|
|
|
|
cases = []
|
|
with open(graphs_path, "r") as f:
|
|
for line in f:
|
|
cases.append(json.loads(line))
|
|
|
|
# Use subset for speed
|
|
cases = cases[:20]
|
|
|
|
# Embed blind
|
|
print(f"Embedding {len(cases)} cases blind of labels...")
|
|
embeddings = []
|
|
case_ids = []
|
|
for c in cases:
|
|
emb = embed_graph(c["graph"], vocab)
|
|
embeddings.append(emb)
|
|
case_ids.append(c["id"])
|
|
|
|
# Align pairs
|
|
print("Computing Procrustes alignment residuals...")
|
|
n = len(cases)
|
|
residuals = np.zeros((n, n))
|
|
|
|
for i in range(n):
|
|
print(f"Aligning case {i}/{n}...", flush=True)
|
|
for j in range(i, n):
|
|
try:
|
|
_, res = conformal_procrustes(embeddings[i], embeddings[j])
|
|
except ValueError:
|
|
res = 1000.0
|
|
except Exception:
|
|
res = 1000.0
|
|
|
|
residuals[i, j] = res
|
|
residuals[j, i] = res
|
|
|
|
# Now load labels and score
|
|
print("Loading labels for evaluation...")
|
|
id_to_label = {}
|
|
with open(labels_path, "r") as f:
|
|
for line in f:
|
|
l = json.loads(line)
|
|
id_to_label[l["id"]] = l["label"]
|
|
|
|
labels = [id_to_label[cid] for cid in case_ids]
|
|
|
|
# Evaluate
|
|
same_res = []
|
|
cross_res = []
|
|
|
|
for i in range(n):
|
|
for j in range(i + 1, n):
|
|
res = residuals[i, j]
|
|
if labels[i] == labels[j]:
|
|
same_res.append(res)
|
|
else:
|
|
cross_res.append(res)
|
|
|
|
same_mean = np.mean(same_res) if same_res else 0.0
|
|
cross_mean = np.mean(cross_res) if cross_res else 0.0
|
|
|
|
print(f"\nSame-structure residual (mean): {same_mean:.6f}")
|
|
print(f"Cross-structure residual (mean): {cross_mean:.6f}")
|
|
print(f"Separability margin: {cross_mean - same_mean:.6f}")
|
|
|
|
if cross_mean > same_mean * 2 and same_mean < 1.0:
|
|
print("\nVERDICT: GO")
|
|
else:
|
|
print("\nVERDICT: NO-GO")
|
|
|
|
if __name__ == "__main__":
|
|
main()
|