fix:批量修复了一些问题

src/llm/eval.py

@@ -25,6 +25,10 @@ async def settle_prediction(prediction_id: int, home_goals: int, away_goals: int
         pred.actual_home_goals = home_goals
         pred.actual_away_goals = away_goals
         pred.settled = True
+        logger.info(
+            "结算完成 prediction_id=%s match=%s actual=%s:%s mode=%s",
+            prediction_id, pred.match_id, home_goals, away_goals, pred.mode or "single",
+        )
         return pred
 
 
@@ -109,8 +113,14 @@ async def get_eval_summary(
         rows = list((await session.execute(stmt)).scalars().all())
 
         from collections import defaultdict
-        buckets: dict[tuple[str, str], dict] = defaultdict(lambda: {
+        buckets: dict[tuple[str, str, str], dict] = defaultdict(lambda: {
             "total": 0, "correct_1x2": 0, "score_errors": [], "conf_sum": 0.0, "conf_count": 0,
+            # 置信度校准分桶(仅 settled 且 pred 完整者计入)
+            "conf_buckets": {
+                "low(0-0.5)": {"total": 0, "correct": 0},
+                "medium(0.5-0.7)": {"total": 0, "correct": 0},
+                "high(0.7-1)": {"total": 0, "correct": 0},
+            },
         })
         evaluated = 0
         skipped_incomplete = 0
@@ -118,35 +128,59 @@ async def get_eval_summary(
             if (p.pred_home_goals is None or p.pred_away_goals is None or p.pred_1x2 is None):
                 skipped_incomplete += 1
                 continue
-            key = (p.provider, p.model)
+            key = (p.provider, p.model, p.prompt_version or "")
             b = buckets[key]
             b["total"] += 1
             evaluated += 1
-            if p.actual_home_goals is None or p.actual_away_goals is None:
-                continue
-            actual = _actual_1x2(p.actual_home_goals, p.actual_away_goals)
-            if p.pred_1x2 == actual:
-                b["correct_1x2"] += 1
-            if p.pred_home_goals is not None and p.pred_away_goals is not None:
+
+            correct = False
+            if p.actual_home_goals is not None and p.actual_away_goals is not None:
+                actual = _actual_1x2(p.actual_home_goals, p.actual_away_goals)
+                if p.pred_1x2 == actual:
+                    b["correct_1x2"] += 1
+                    correct = True
+            if (
+                p.pred_home_goals is not None and p.pred_away_goals is not None
+                and p.actual_home_goals is not None and p.actual_away_goals is not None
+            ):
                 err = ((p.pred_home_goals - p.actual_home_goals) ** 2 +
                        (p.pred_away_goals - p.actual_away_goals) ** 2) ** 0.5
                 b["score_errors"].append(err)
             if p.subjective_confidence is not None:
                 b["conf_sum"] += p.subjective_confidence
                 b["conf_count"] += 1
+                # 仅当有实际结果可用于校准时,才落入置信度分桶
+                if p.actual_home_goals is not None and p.actual_away_goals is not None:
+                    conf = p.subjective_confidence
+                    if conf < 0.5:
+                        bucket = "low(0-0.5)"
+                    elif conf < 0.7:
+                        bucket = "medium(0.5-0.7)"
+                    else:
+                        bucket = "high(0.7-1)"
+                    b["conf_buckets"][bucket]["total"] += 1
+                    if correct:
+                        b["conf_buckets"][bucket]["correct"] += 1
 
         summary = []
-        for (prov, model), b in sorted(buckets.items()):
+        for (prov, model, ver), b in sorted(buckets.items()):
             acc = (b["correct_1x2"] / b["total"] * 100) if b["total"] else 0
             avg_err = (sum(b["score_errors"]) / len(b["score_errors"])) if b["score_errors"] else None
             avg_conf = (b["conf_sum"] / b["conf_count"]) if b["conf_count"] else None
+            # 校准分桶 → 命中率
+            calibration = {}
+            for name, cb in b["conf_buckets"].items():
+                hit_rate = round(cb["correct"] / cb["total"] * 100, 1) if cb["total"] else None
+                calibration[name] = {"total": cb["total"], "hit_rate": hit_rate}
             summary.append({
                 "provider": prov,
                 "model": model,
+                "prompt_version": ver or None,
                 "total": b["total"],
                 "accuracy_1x2": round(acc, 1),
                 "avg_score_rmse": round(avg_err, 2) if avg_err is not None else None,
                 "avg_subjective_confidence": round(avg_conf, 2) if avg_conf is not None else None,
+                "calibration": calibration,
             })
         return {
             "summary": summary,