fix(lotto): 학습 게이트 정직화 (engine-best vs random-best 6trial·명시적 gated·정체성 일관)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

lotto/app/backtest.py

@@ -6,6 +6,9 @@ from typing import Any, Dict, List, Optional, Tuple
 from .analyzer import build_analysis_cache, build_number_weights, score_combination
 from .utils import weighted_sample_6
 
+# engine_w trials 수와 동일하게 맞춰 selection bias를 상쇄한다.
+N_NULL_TRIALS = 6
+
 
 def grade_tickets(tickets: List[List[int]], winning6: List[int], bonus: int) -> Dict[str, Any]:
     """티켓 묶음을 당첨번호로 채점 → 매칭 히스토그램 + 보너스 + best_match.
@@ -194,8 +197,10 @@ def run_forward_purchase(draw_no: int, k: int = 5000, pool_n: int = 20000,
         bought = purchase_tickets(pool, cache, base, k)
         _store("engine_w", "base", base, None, bought)
 
-    # 2) random_null
+    # 2) random_null — N_NULL_TRIALS 개 (engine_w 수와 동일해 selection bias 상쇄)
-    _store("random_null", "-", None, None, random_null_tickets(k, seed=sample_seed))
+    for _i in range(N_NULL_TRIALS):
+        seed_i = None if sample_seed is None else sample_seed + 100 + _i
+        _store("random_null", f"r{_i}", None, None, random_null_tickets(k, seed=seed_i))
     # 3) coverage
     _store("coverage", "-", None, None, coverage_tickets(k, seed=sample_seed))
 

lotto/app/weight_evolver.py

@@ -19,7 +19,9 @@ DEFAULT_UNIFORM = [0.2] * N_METRICS  # cold start
 RANK_BY_CORRECT = {6: 1, 5: 3, 4: 4, 3: 5}
 RANK_BONUS = {1: 1.0, 2: 0.8, 3: 0.6, 4: 0.3, 5: 0.1}
 
-LIFT_EPSILON = 0.5  # 등수점수 노이즈 게이팅 임계 (튜닝 가능)
+LIFT_EPSILON = 10.0  # best-of-engine vs best-of-random margin;
+# selection bias already cancelled by equal group sizes (N_NULL_TRIALS == engine trial count);
+# tune as needed.
 
 PRIZE_WEIGHTS = {"m6": 1000.0, "bonus_hits": 50.0, "m5": 30.0, "m4": 4.0, "m3": 1.0}
 
@@ -35,7 +37,9 @@ def select_winner_by_lift(per_w: List[Dict[str, Any]], random_score: float,
 
 def prize_score_from_hist(hist: Dict[str, int]) -> float:
     """매칭 히스토그램 → 등수 가중 합산 점수.
-    1등=m6, 2등=bonus_hits, 3등=m5−bonus_hits, 4등=m4, 5등=m3."""
+    1등=m6, 2등=bonus_hits, 3등=m5−bonus_hits, 4등=m4, 5등=m3.
+    m3/m4/m5/m6/bonus_hits 키만 읽으며 나머지는 무시하므로
+    DB 전체 행(backtest_runs row)을 그대로 넘겨도 안전하다."""
     third = max(0, hist.get("m5", 0) - hist.get("bonus_hits", 0))
     return (hist.get("m6", 0) * PRIZE_WEIGHTS["m6"]
             + hist.get("bonus_hits", 0) * PRIZE_WEIGHTS["bonus_hits"]
@@ -294,34 +298,46 @@ def evaluate_weekly() -> Dict[str, Any]:
 
     winner = max(per_day, key=lambda d: d["avg_score"])
 
-    # 자가학습 강화: backtest forward 등수점수 lift로 winner 재선정
+    # 자가학습 강화: backtest forward 등수점수 lift로 winner 재선정.
+    # best-of-engine vs best-of-random 비교 — 동등 그룹 크기로 selection bias 상쇄.
     latest_no = latest["drw_no"]
     runs = db.get_backtest_runs(draw_no=latest_no)
     engine_runs = [r for r in runs if r["strategy"] == "engine_w"]
     null_runs = [r for r in runs if r["strategy"] == "random_null"]
+    gated = False  # 이후 decide_base_update override에 사용
     if engine_runs and null_runs:
-        random_score = prize_score_from_hist(null_runs[0])
+        # base 단독 행이 있고 w* 행도 있으면 base 행 제외 (identity collision 방지)
+        has_w_trials = any(r["weight_label"].startswith("w") for r in engine_runs)
+        if has_w_trials:
+            engine_runs = [r for r in engine_runs if r["weight_label"] != "base"]
+
+        # best-of-random: 동등 그룹의 최댓값 (selection bias 상쇄)
+        random_best = max(prize_score_from_hist(r) for r in null_runs)
+
         per_w = []
         for r in engine_runs:
             per_w.append({
                 "trial_id": r["trial_id"],
-                "day_of_week": int(r["weight_label"][1:]) if r["weight_label"].startswith("w") else 0,
+                "weight_label": r["weight_label"],
                 "weight": json.loads(r["weight_json"]) if r["weight_json"] else DEFAULT_UNIFORM[:],
                 "prize_score": prize_score_from_hist(r),
+                "best_match": r["best_match"],
             })
-        lift_winner = select_winner_by_lift(per_w, random_score=random_score)
+
+        lift_winner = select_winner_by_lift(per_w, random_score=random_best)
         if not lift_winner["gated"]:
+            # lift winner의 정체성과 채점값을 일관되게 사용
             winner = {
                 "trial_id": lift_winner["trial_id"],
-                "day_of_week": lift_winner["day_of_week"],
                 "weight": lift_winner["weight"],
-                "avg_score": winner["avg_score"],
+                "max_correct": lift_winner["best_match"],   # 이 trial의 실제값
-                "max_correct": winner["max_correct"],
+                "avg_score": lift_winner["prize_score"],    # lift winner의 prize score
                 "lift": lift_winner["lift"],
             }
         else:
-            # 노이즈 → base 유지 강제 (max_correct를 0으로 낮춰 unchanged 유도)
+            # 노이즈 → gated 플래그 설정; decide_base_update 이후 명시적으로 override
-            winner = {**winner, "max_correct": min(winner["max_correct"], 2), "lift": lift_winner["lift"]}
+            gated = True
+            winner = {**winner, "lift": lift_winner["lift"]}
 
     current_base = db.get_current_base()
     new_base, reason = decide_base_update(
@@ -330,6 +346,11 @@ def evaluate_weekly() -> Dict[str, Any]:
         current_base=current_base,
     )
 
+    # gated path: decide_base_update 결과와 무관하게 base 유지 강제
+    if gated:
+        new_base = list(current_base) if current_base is not None else DEFAULT_UNIFORM[:]
+        reason = "unchanged_gated"
+
     next_monday = today + timedelta(days=(7 - today.weekday()) % 7 or 7)
     next_monday_iso = next_monday.isoformat()
 

lotto/tests/test_backtest_db.py

@@ -182,6 +182,7 @@ def test_track_record_and_review_payload(monkeypatch):
 
     tr = bt.track_record()
     assert "random_null" in tr["by_strategy"]
+    # 이제 random_null은 N_NULL_TRIALS=6 행이므로 6*20=120장
     assert tr["by_strategy"]["random_null"]["n_tickets"] >= 20
 
     payload = bt.build_review_payload(40)
@@ -191,3 +192,129 @@ def test_track_record_and_review_payload(monkeypatch):
     assert "calibration_trend" in payload
     assert payload["winner_analysis"] is not None
     assert "score_total" in payload["winner_analysis"]
+
+
+def test_run_forward_purchase_random_null_count(monkeypatch):
+    """run_forward_purchase는 random_null을 N_NULL_TRIALS=6개 저장해야 한다."""
+    db = _fresh_db(monkeypatch)
+    _seed_draws(db, 40)
+    from app import backtest as bt
+    res = bt.run_forward_purchase(draw_no=40, k=20, pool_n=500, sample_seed=7)
+    assert res["ok"] is True
+    rows = db.get_backtest_runs(draw_no=40)
+    null_rows = [r for r in rows if r["strategy"] == "random_null"]
+    assert len(null_rows) == bt.N_NULL_TRIALS  # 6개
+    null_labels = {r["weight_label"] for r in null_rows}
+    assert null_labels == {f"r{i}" for i in range(bt.N_NULL_TRIALS)}
+    for r in null_rows:
+        assert r["n_tickets"] == 20
+
+
+def test_evaluate_weekly_gated_keeps_base_unchanged(monkeypatch):
+    """Fix 5 통합 테스트 (end-to-end gated path).
+
+    접근: DB에 draws, weight_trials, auto_picks, backtest_runs, base_history를 직접 심어
+    evaluate_weekly()의 gated 분기가 base를 바꾸지 않음을 검증한다.
+
+    gated 조건: engine_w 최고 prize_score − random_best < LIFT_EPSILON(10.0).
+    engine_best=5, random_best=20 → lift=-15 → gated.
+
+    evaluate_weekly 내부 흐름:
+      - get_weekly_trials(week_start) : _today_kst() 기준 week_start 사용
+      - get_latest_draw()             : draws 테이블에서 max(drw_no) 반환
+    두 참조가 같은 날짜 기준이어야 하므로 _today_kst를 monkeypatch로 고정하고
+    draws의 최신 회차 날짜(drw_date)를 해당 주의 날짜로 맞춘다.
+    """
+    import json as _json
+    from datetime import date, timedelta, datetime as _dt, timezone as _tz, timedelta as _td
+
+    db = _fresh_db(monkeypatch)
+
+    # KST 오늘 날짜 — evaluate_weekly가 이 날짜를 기준으로 week_start 계산
+    KST = _tz(_td(hours=9))
+    today_kst = _dt.now(KST).date()
+    from app import weight_evolver as we
+    week_start = we.get_week_start(today_kst)
+
+    # 1) draws 심기 — 최신 회차의 drw_date를 week_start 주 안의 날짜로 맞춤
+    import random as _r; _r.seed(99)
+    rows = []
+    for i in range(1, 41):
+        s = sorted(_r.sample(range(1, 46), 6))
+        # 마지막 회차(40)는 오늘 날짜 사용 (week_start 주 내)
+        if i == 40:
+            drw_date = today_kst.isoformat()
+        else:
+            drw_date = f"2020-01-{(i % 28) + 1:02d}"
+        rows.append({
+            "drw_no": i, "drw_date": drw_date,
+            "n1": s[0], "n2": s[1], "n3": s[2],
+            "n4": s[3], "n5": s[4], "n6": s[5],
+            "bonus": (s[5] % 45) + 1,
+        })
+    db.upsert_many_draws(rows)
+    latest = db.get_latest_draw()
+    assert latest is not None
+    assert latest["drw_date"] == today_kst.isoformat()
+
+    # 2) weight trial 1개 심기 (day_of_week=0, week_start=오늘 주)
+    trial_w = [0.2, 0.2, 0.2, 0.2, 0.2]
+    db.save_weight_trial(week_start, 0, trial_w, "perturb")
+    trial_rows = db.get_weekly_trials(week_start)
+    assert len(trial_rows) == 1
+    trial_id = trial_rows[0]["id"]
+
+    # 3) auto_picks 1개 심기 (winning 번호와 2개 일치 → max_correct=2)
+    winning6 = [latest["n1"], latest["n2"], latest["n3"],
+                latest["n4"], latest["n5"], latest["n6"]]
+    pick = winning6[:2] + [40, 41, 42, 43]
+    db.save_auto_pick(trial_id, 1, pick, meta_score=0.5)
+
+    # 4) backtest_runs: engine_w prize_score=5, random_null 6개 prize_score=20 (gated 확실)
+    LOW_HIST = {"m3": 5, "m4": 0, "m5": 0, "m6": 0, "bonus_hits": 0}   # prize=5
+    HIGH_HIST = {"m3": 20, "m4": 0, "m5": 0, "m6": 0, "bonus_hits": 0}  # prize=20
+    draw_no = latest["drw_no"]
+    db.save_backtest_run(
+        draw_no=draw_no, strategy="engine_w", weight_label="w0",
+        weight_json=_json.dumps(trial_w), trial_id=trial_id, n_tickets=20,
+        hist=LOW_HIST, best_match=2, avg_meta_score=0.5,
+    )
+    from app import backtest as bt
+    for i in range(bt.N_NULL_TRIALS):
+        db.save_backtest_run(
+            draw_no=draw_no, strategy="random_null", weight_label=f"r{i}",
+            weight_json=None, trial_id=None, n_tickets=20,
+            hist=HIGH_HIST, best_match=3, avg_meta_score=0.5,
+        )
+
+    # 5) current base 저장 (이전 주 월요일 effective_from)
+    base_w = [0.2, 0.2, 0.2, 0.2, 0.2]
+    prev_monday = (today_kst - timedelta(weeks=1, days=today_kst.weekday())).isoformat()
+    db.save_base_history(
+        effective_from=prev_monday,
+        weight=base_w,
+        source_trial_id=None,
+        update_reason="cold_start",
+        winner_score=None,
+        winner_max_correct=None,
+    )
+    assert db.get_current_base() == base_w
+
+    # 6) evaluate_weekly 호출 — _today_kst()를 monkeypatch로 오늘 날짜 고정
+    monkeypatch.setattr(we, "_today_kst", lambda: today_kst)
+
+    result = we.evaluate_weekly()
+
+    assert result.get("ok") is True, f"evaluate_weekly 실패: {result}"
+
+    # gated path 검증
+    update_reason = result.get("update_reason", "")
+    assert update_reason in ("unchanged_gated", "idempotent_skip"), (
+        f"gated여야 하는데 reason='{update_reason}' — 게이팅 로직 깨짐"
+    )
+
+    # base가 바뀌지 않았는지 검증
+    new_base = result.get("new_base")
+    assert new_base == base_w, (
+        f"gated인데 base가 변경됨: {new_base} != {base_w}"
+    )

lotto/tests/test_weight_evolver.py

@@ -129,7 +129,7 @@ def test_select_winner_by_lift_gating():
         {"trial_id": 2, "day_of_week": 1, "weight": [0.3,0.2,0.2,0.2,0.1], "prize_score": 9.0},
         {"trial_id": 3, "day_of_week": 2, "weight": [0.1,0.3,0.2,0.2,0.2], "prize_score": 4.0},
     ]
-    # random baseline이 8.0이면 lift는 +1, +1, -4 → 노이즈 ε=2 안에서 게이팅
+    # random baseline이 8.0이면 lift는 -3, +1, -4 → 최대 lift(+1) < ε(2) → 게이팅
     winner = we.select_winner_by_lift(per_w, random_score=8.0, epsilon=2.0)
     assert winner["gated"] is True       # 최대 lift(+1) < ε(2) → 게이팅
     winner2 = we.select_winner_by_lift(per_w, random_score=3.0, epsilon=2.0)
@@ -142,3 +142,57 @@ def test_prize_score_from_hist():
     s = we.prize_score_from_hist({"m3": 10, "m4": 2, "m5": 0, "m6": 0, "bonus_hits": 0})
     s_big = we.prize_score_from_hist({"m3": 0, "m4": 0, "m5": 0, "m6": 1, "bonus_hits": 0})
     assert s_big > s   # 1등 1장이 5등 다수보다 큼
+
+
+def test_select_winner_by_lift_preserves_all_keys():
+    """select_winner_by_lift는 per_w 항목의 모든 키를 보존해야 한다.
+    best_match, weight_label 등 identity 필드가 누락되면 evaluate_weekly가 깨진다."""
+    per_w = [
+        {
+            "trial_id": 10,
+            "weight_label": "w0",
+            "weight": [0.2] * 5,
+            "prize_score": 3.0,
+            "best_match": 3,
+        },
+        {
+            "trial_id": 11,
+            "weight_label": "w1",
+            "weight": [0.3, 0.2, 0.2, 0.2, 0.1],
+            "prize_score": 20.0,
+            "best_match": 4,
+        },
+    ]
+    result = we.select_winner_by_lift(per_w, random_score=5.0, epsilon=2.0)
+    assert result["gated"] is False
+    assert result["trial_id"] == 11
+    assert result["weight_label"] == "w1"   # identity 키 보존
+    assert result["best_match"] == 4        # best_match 키 보존
+    assert "lift" in result                 # lift 추가됨
+    assert result["lift"] == pytest.approx(15.0)
+
+
+def test_gated_path_keeps_base_via_select_winner():
+    """gated=True일 때 select_winner_by_lift의 반환값 검증.
+    evaluate_weekly 내의 gated 분기가 올바른 값에 의존함을 확인한다."""
+    per_w = [
+        {"trial_id": 1, "weight_label": "w0", "weight": [0.2]*5,
+         "prize_score": 5.0, "best_match": 2},
+        {"trial_id": 2, "weight_label": "w1", "weight": [0.3,0.2,0.2,0.2,0.1],
+         "prize_score": 7.0, "best_match": 3},
+    ]
+    # random_best=8.0 → 최대 engine lift=7-8=-1 → gated
+    result = we.select_winner_by_lift(per_w, random_score=8.0, epsilon=we.LIFT_EPSILON)
+    assert result["gated"] is True
+    assert result["lift"] < 0
+
+    # decide_base_update를 통해 gated가 unchanged를 유도하는지 확인
+    # (gated override가 없더라도, 현재 LIFT_EPSILON=10.0 하에서 lift<0이면 항상 gated)
+    current = [0.2, 0.2, 0.2, 0.2, 0.2]
+    # gated이면 evaluate_weekly가 current_base를 그대로 유지해야 함
+    # 여기서는 override 로직을 직접 재현해 검증한다
+    gated = result["gated"]
+    new_base_override = list(current) if gated else None
+    reason_override = "unchanged_gated" if gated else "should_not_reach"
+    assert new_base_override == current
+    assert reason_override == "unchanged_gated"