ML Pipeline

Enabled by default (modules.ml = true). Detects ML anti-patterns that cause silent bugs or irreproducible results. Prefix: ML-

Finding	Severity	What it matches
ML-001	Error	fit() called on test data (data leakage)
ML-002	Warning	train_test_split without random_state
ML-003	Warning	pickle for model serialization
ML-004	Warning	Hardcoded absolute data file paths
ML-005	Warning	model.fit() called inside a loop (repeated fitting)
ML-006	Warning	torch.no_grad() without model.eval()
ML-007	Warning	Hardcoded absolute/home path in .to_csv()/.to_parquet()
ML-008	Info	Random operations without seed (non-reproducible)
ML-009	Info	train_test_split without stratify
ML-010	Info	Deprecated sklearn import paths

# Bad — data leakage
scaler.fit(X_test)

# Good
scaler.fit(X_train)
scaler.transform(X_test)

# Bad — non-reproducible
X_train, X_test = train_test_split(X, y)

# Good
X_train, X_test = train_test_split(X, y, random_state=42)

# Bad — repeated fitting in a loop
for epoch in range(100):
    model.fit(X_train, y_train)  # refits from scratch every iteration

# Good — train once, or use partial_fit() for incremental learning
model.fit(X_train, y_train)

# Bad — missing model.eval() before inference
with torch.no_grad():
    outputs = model(inputs)

# Good
model.eval()
with torch.no_grad():
    outputs = model(inputs)

# Bad — random operations without seed
X = np.random.randn(100, 10)

# Good
np.random.seed(42)
X = np.random.randn(100, 10)