diff --git a/scripts/train_stage1.sh b/scripts/train_stage1.sh
index a8e6567..819db4a 100755
--- a/scripts/train_stage1.sh
+++ b/scripts/train_stage1.sh
@@ -8,6 +8,7 @@ MAX_STEPS="${2:-10000}"
 CKPT_DIR="${3:-checkpoints/stage1}"
 
 cd "$(dirname "$0")/.."
+export PYTHONPATH=src PYTHONUNBUFFERED=1
 .venv/bin/python -m tsmm.train.stage1 \
   --data "$DATA" \
   --ckpt_dir "$CKPT_DIR" \
diff --git a/scripts/train_stage2.sh b/scripts/train_stage2.sh
index 55c57c8..d38f3c3 100755
--- a/scripts/train_stage2.sh
+++ b/scripts/train_stage2.sh
@@ -8,6 +8,7 @@ STAGE1_CKPT="${3:-checkpoints/stage1/stage1_final.pt}"
 CKPT_DIR="${4:-checkpoints/stage2}"
 
 cd "$(dirname "$0")/.."
+export PYTHONPATH=src PYTHONUNBUFFERED=1
 .venv/bin/python -m tsmm.train.stage2 \
   --data "$DATA" \
   --stage1_ckpt "$STAGE1_CKPT" \
diff --git a/tests/test_instruct_check.py b/tests/test_instruct_check.py
new file mode 100644
index 0000000..41d76b9
--- /dev/null
+++ b/tests/test_instruct_check.py
@@ -0,0 +1,91 @@
+"""Unit tests for eval.instruct_check (T4.2) — CPU helpers only.
+
+The model-driven ``evaluate`` path is exercised live in M4 exit validation.
+"""
+import json
+import os
+import tempfile
+
+import pytest
+
+from tsmm.eval.instruct_check import (
+    CATEGORIES,
+    _forecast_list,
+    load_by_category,
+)
+
+
+def _write_eval(tmpdir, rows):
+    path = os.path.join(tmpdir, "eval.jsonl")
+    with open(path, "w") as f:
+        for r in rows:
+            f.write(json.dumps(r) + "\n")
+    return path
+
+
+def test_load_by_category_buckets_and_caps(tmp_path):
+    rows = []
+    for cat in CATEGORIES:
+        for i in range(5):
+            rows.append({"category": cat, "series": [[1.0]], "answer": "x", "labels": [0]})
+    # plus an unknown-category row that must be ignored
+    rows.append({"category": "mystery", "series": [[1.0]], "answer": "x"})
+    path = _write_eval(str(tmp_path), rows)
+    out = load_by_category(path, per_cat=3, seed=0)
+    assert set(out.keys()) == set(CATEGORIES)
+    for cat in CATEGORIES:
+        assert len(out[cat]) == 3
+    assert "mystery" not in out
+
+
+def test_load_by_category_deterministic_with_seed(tmp_path):
+    rows = [{"category": "anomaly", "series": [[float(i)]], "answer": str(i), "labels": [0]}
+            for i in range(20)]
+    path = _write_eval(str(tmp_path), rows)
+    a = load_by_category(path, per_cat=5, seed=42)
+    b = load_by_category(path, per_cat=5, seed=42)
+    assert [s["answer"] for s in a["anomaly"]] == [s["answer"] for s in b["anomaly"]]
+    # different seed → (very likely) different selection
+    c = load_by_category(path, per_cat=5, seed=7)
+    assert [s["answer"] for s in a["anomaly"]] != [s["answer"] for s in c["anomaly"]]
+
+
+def test_load_by_category_handles_short_pool(tmp_path):
+    rows = [{"category": "event", "series": [[1.0]], "answer": "x"} for _ in range(2)]
+    path = _write_eval(str(tmp_path), rows)
+    out = load_by_category(path, per_cat=10, seed=0)
+    assert len(out["event"]) == 2
+    # missing categories are empty lists, not errors
+    assert out["anomaly"] == []
+
+
+def test_load_by_category_skips_blank_lines(tmp_path):
+    path = os.path.join(str(tmp_path), "eval.jsonl")
+    with open(path, "w") as f:
+        f.write(json.dumps({"category": "anomaly", "series": [[1.0]], "answer": "x"}) + "\n")
+        f.write("\n")
+        f.write("   \n")
+        f.write(json.dumps({"category": "anomaly", "series": [[2.0]], "answer": "y"}) + "\n")
+    out = load_by_category(path, per_cat=5, seed=0)
+    assert len(out["anomaly"]) == 2
+
+
+def test_forecast_list_from_nested_json():
+    ref = json.dumps({"forecast": [[1.0, 2.0, 3.0]]})
+    assert _forecast_list(ref) == [1.0, 2.0, 3.0]
+
+
+def test_forecast_list_from_flat_json():
+    ref = json.dumps({"forecast": [4.0, 5.0]})
+    assert _forecast_list(ref) == [4.0, 5.0]
+
+
+def test_forecast_list_accepts_dict_input():
+    assert _forecast_list({"forecast": [[1.0, 2.0]]}) == [1.0, 2.0]
+
+
+def test_forecast_list_returns_empty_on_garbage():
+    assert _forecast_list("not json") == []
+    assert _forecast_list(json.dumps({"no_forecast": 1})) == []
+    assert _forecast_list("") == []
+    assert _forecast_list(None) == []