Deterministic fake dates; refresh data and scripts

data/test.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c3d4d7f430baf6346c9fe85220b9f79de23da0ced00b1ab78a2524f762798b91
-size 538322
+oid sha256:e7a6d000be881bf3c4367ebe4592e529afb943e1509e780ca7785b8013ba013a
+size 538386

data/train.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8e41f9a07533a221aa4c9d017fc314ac41d16bbda635d0e58750b585010b0b77
-size 276923
+oid sha256:a065c066beccba56f6da7bc4d8a5924024937d3740edac2bb7400acbe9800954
+size 277000

full/dataset_v2.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5e67543f91b473472150ccf0f83ada820b0282ccbf8aaead85e6e0e0abc01568
-size 790168
+oid sha256:9ccedfe50f9bbb11e3cf1b92f7775b75691f07945ed06f8744787e0766384212
+size 789975

scripts/build_dataset.py

@@ -1,10 +1,11 @@
 """Verified synthetic rows -> deduped, balanced, PII-filled dataset with a stratified 1/3 train / 2/3 test split
 in which near-duplicate clusters never cross the split."""
-import argparse, re, random, unicodedata
+import argparse, datetime, re, random, unicodedata
 import numpy as np, pandas as pd
 from faker import Faker
 from sentence_transformers import SentenceTransformer
 
+END = datetime.date(2026, 1, 1)   # fixed so fake dates do not depend on the clock
 TYPES = {"NAME": "name", "AMOUNT": "amount", "LAST4": "card_digits", "MERCHANT": "merchant", "CITY": "location", "DATE": "date", "REF": "reference"}
 MERCHANTS = ["Amazon", "Starbucks", "Tesco", "Uber", "Netflix", "Spotify", "Walmart", "Target", "IKEA", "Zara", "Airbnb", "Apple Store",
              "Best Buy", "Lidl", "Deliveroo", "Ryanair", "Booking.com", "Nike", "eBay", "Shell"]
@@ -17,7 +18,7 @@ def fill(t, fk, rng):
     gen = {"NAME": lambda: fk.first_name() if rng.random() < 0.6 else fk.name(),
            "AMOUNT": lambda: rng.choice(["£", "£", "£", "$", "€", ""]) + str(rng.choice([rng.randint(5, 99), rng.randint(100, 999), round(rng.uniform(5, 900), 2)])),
            "LAST4": lambda: f"{rng.randint(0, 9999):04d}", "MERCHANT": lambda: rng.choice(MERCHANTS), "CITY": fk.city,
-           "DATE": lambda: rng.choice([fk.date_object().strftime("%d/%m"), fk.date_object().strftime("%-d %B"), f"the {rng.randint(1, 28)}th"]),
+           "DATE": lambda: rng.choice([fk.date_object(END).strftime("%d/%m"), fk.date_object(END).strftime("%-d %B"), f"the {rng.randint(1, 28)}th"]),
            "REF": lambda: rng.choice([f"TXN-{rng.randint(100000, 999999)}", f"{rng.randint(10**7, 10**8 - 1)}", f"REF{rng.randint(1000, 9999)}X"])}
     return re.sub(r"\{(\w+)\}", lambda m: str(gen[m.group(1)]()), t)
 

scripts/verify_seeds.py

@@ -80,7 +80,13 @@ def main():
     for i in np.random.default_rng(1).permutation(np.where(prev & ~(lr_bad | knn_bad))[0])[:10]:
         print(f"  {saved.text[i]!r}  labeled={names[i]}  model says={cls[p_lr[i].argmax()]}")
     out = saved.assign(lr_bad=lr_bad, knn_bad=knn_bad, both_bad=both)
+    out["votes"] = out.suspect.astype(int) + out.lr_bad.astype(int) + out.knn_bad.astype(int)
+    out["suspect_consensus"] = out.votes >= 2                                # flagged by >= 2 of 3 detectors
     out.to_parquet("data/seeds_audited.parquet")
+    # artifacts used by select_clean.py (confusion graph over the same embeddings)
+    np.save("data/seed_emb_bge.npy", emb)
+    np.save("data/seed_cvprob_bge.npy", cross_val_predict(LogisticRegression(max_iter=3000, C=20), emb, y,
+                                                          cv=StratifiedKFold(5, shuffle=True, random_state=7), method="predict_proba"))
 
 if __name__ == "__main__":
     main()