From 3196552c781a5f7ac144c89a63cfb6b16377167c Mon Sep 17 00:00:00 2001
From: "2569718930@qq.com" <2569718930@qq.com>
Date: Fri, 20 Mar 2026 21:30:52 +0800
Subject: [PATCH] Archive probability snapshots and wire them into training

---
 .../probability_calibration/default.json      |   4 +-
 .../evaluation_report.json                    | 120 +++----
 .../training_samples.json                     | 317 +++++++++++------
 data/probability_training_snapshots.jsonl     |   1 +
 docs/PROBABILITY_SNAPSHOT_ARCHIVE_ZH.md       | 324 ++++++++++++++++++
 .../export_probability_training_dataset.py    |  15 +-
 scripts/fit_probability_calibration.py        | 149 +++++++-
 src/analysis/probability_snapshot_archive.py  | 187 ++++++++++
 src/analysis/trend_engine.py                  |  21 ++
 tests/test_probability_snapshot_archive.py    | 117 +++++++
 tests/test_probability_training_dataset.py    |  45 +++
 11 files changed, 1129 insertions(+), 171 deletions(-)
 create mode 100644 data/probability_training_snapshots.jsonl
 create mode 100644 docs/PROBABILITY_SNAPSHOT_ARCHIVE_ZH.md
 create mode 100644 src/analysis/probability_snapshot_archive.py
 create mode 100644 tests/test_probability_snapshot_archive.py
 create mode 100644 tests/test_probability_training_dataset.py

diff --git a/artifacts/probability_calibration/default.json b/artifacts/probability_calibration/default.json
index 80ebc4b4..281fbf86 100644
--- a/artifacts/probability_calibration/default.json
+++ b/artifacts/probability_calibration/default.json
@@ -1,6 +1,6 @@
 {
-  "version": "emos-20260320130245",
-  "trained_at": "2026-03-20T13:02:45.903772+00:00",
+  "version": "emos-20260320132525",
+  "trained_at": "2026-03-20T13:25:25.836021+00:00",
   "global": {
     "mu": {
       "intercept": -1.57406048,
diff --git a/artifacts/probability_calibration/evaluation_report.json b/artifacts/probability_calibration/evaluation_report.json
index 9bb1db0c..f3b787d1 100644
--- a/artifacts/probability_calibration/evaluation_report.json
+++ b/artifacts/probability_calibration/evaluation_report.json
@@ -8,239 +8,239 @@
       "bucket_hit_rate": 0.695238
     },
     "emos": {
-      "mean_crps": 2.650216,
-      "mean_mae": 2.722829,
-      "bucket_hit_rate": 0.666667
+      "mean_crps": 2.700275,
+      "mean_mae": 2.721143,
+      "bucket_hit_rate": 0.695238
     },
     "delta": {
-      "crps": -0.143722,
-      "mae": 0.001686,
-      "bucket_hit_rate": -0.028571
+      "crps": -0.093663,
+      "mae": 0.0,
+      "bucket_hit_rate": 0.0
     }
   },
   "by_city": {
     "ankara": {
       "samples": 7,
       "legacy_mean_crps": 2.023242,
-      "emos_mean_crps": 2.015458,
+      "emos_mean_crps": 2.023242,
       "legacy_mean_mae": 1.984286,
-      "emos_mean_mae": 1.977721,
+      "emos_mean_mae": 1.984286,
       "legacy_bucket_hit_rate": 0.714286,
       "emos_bucket_hit_rate": 0.714286
     },
     "atlanta": {
       "samples": 5,
       "legacy_mean_crps": 12.792034,
-      "emos_mean_crps": 12.526108,
+      "emos_mean_crps": 12.694543,
       "legacy_mean_mae": 12.806,
-      "emos_mean_mae": 12.919677,
+      "emos_mean_mae": 12.806,
       "legacy_bucket_hit_rate": 0.6,
-      "emos_bucket_hit_rate": 0.4
+      "emos_bucket_hit_rate": 0.6
     },
     "buenos aires": {
       "samples": 5,
       "legacy_mean_crps": 3.846144,
-      "emos_mean_crps": 3.743967,
+      "emos_mean_crps": 3.846144,
       "legacy_mean_mae": 4.168,
-      "emos_mean_mae": 4.086663,
+      "emos_mean_mae": 4.168,
       "legacy_bucket_hit_rate": 0.6,
       "emos_bucket_hit_rate": 0.6
     },
     "chicago": {
       "samples": 4,
       "legacy_mean_crps": 1.346667,
-      "emos_mean_crps": 0.603128,
+      "emos_mean_crps": 0.601765,
       "legacy_mean_mae": 0.0,
-      "emos_mean_mae": 0.075265,
+      "emos_mean_mae": 0.0,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 1.0
     },
     "dallas": {
       "samples": 4,
       "legacy_mean_crps": 1.256111,
-      "emos_mean_crps": 0.65514,
+      "emos_mean_crps": 0.651425,
       "legacy_mean_mae": 0.0,
-      "emos_mean_mae": 0.114807,
+      "emos_mean_mae": 0.0,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 1.0
     },
     "hong kong": {
       "samples": 3,
       "legacy_mean_crps": 0.261027,
-      "emos_mean_crps": 0.269856,
+      "emos_mean_crps": 0.261027,
       "legacy_mean_mae": 0.1,
-      "emos_mean_mae": 0.174879,
+      "emos_mean_mae": 0.1,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 0.666667
     },
     "london": {
       "samples": 6,
       "legacy_mean_crps": 2.079624,
-      "emos_mean_crps": 2.005311,
+      "emos_mean_crps": 2.079624,
       "legacy_mean_mae": 2.451667,
-      "emos_mean_mae": 2.327304,
+      "emos_mean_mae": 2.451667,
       "legacy_bucket_hit_rate": 0.166667,
-      "emos_bucket_hit_rate": 0.333333
+      "emos_bucket_hit_rate": 0.166667
     },
     "lucknow": {
       "samples": 4,
       "legacy_mean_crps": 1.468528,
-      "emos_mean_crps": 1.443159,
+      "emos_mean_crps": 1.468528,
       "legacy_mean_mae": 1.6025,
-      "emos_mean_mae": 1.682697,
+      "emos_mean_mae": 1.6025,
       "legacy_bucket_hit_rate": 0.5,
       "emos_bucket_hit_rate": 0.5
     },
     "madrid": {
       "samples": 2,
       "legacy_mean_crps": 6.27726,
-      "emos_mean_crps": 6.090563,
+      "emos_mean_crps": 6.27726,
       "legacy_mean_mae": 7.33,
-      "emos_mean_mae": 7.141649,
+      "emos_mean_mae": 7.33,
       "legacy_bucket_hit_rate": 0.0,
       "emos_bucket_hit_rate": 0.0
     },
     "miami": {
       "samples": 5,
       "legacy_mean_crps": 11.665378,
-      "emos_mean_crps": 11.493094,
+      "emos_mean_crps": 11.665378,
       "legacy_mean_mae": 12.07,
-      "emos_mean_mae": 12.157623,
+      "emos_mean_mae": 12.07,
       "legacy_bucket_hit_rate": 0.6,
-      "emos_bucket_hit_rate": 0.2
+      "emos_bucket_hit_rate": 0.6
     },
     "milan": {
       "samples": 3,
       "legacy_mean_crps": 4.401392,
-      "emos_mean_crps": 3.837715,
+      "emos_mean_crps": 3.928883,
       "legacy_mean_mae": 4.06,
-      "emos_mean_mae": 3.942288,
+      "emos_mean_mae": 4.06,
       "legacy_bucket_hit_rate": 0.666667,
       "emos_bucket_hit_rate": 0.666667
     },
     "munich": {
       "samples": 6,
       "legacy_mean_crps": 2.988583,
-      "emos_mean_crps": 2.918299,
+      "emos_mean_crps": 2.988583,
       "legacy_mean_mae": 3.143333,
-      "emos_mean_mae": 3.126482,
+      "emos_mean_mae": 3.143333,
       "legacy_bucket_hit_rate": 0.5,
       "emos_bucket_hit_rate": 0.5
     },
     "new york": {
       "samples": 4,
       "legacy_mean_crps": 1.861101,
-      "emos_mean_crps": 1.363203,
+      "emos_mean_crps": 1.409393,
       "legacy_mean_mae": 1.3725,
-      "emos_mean_mae": 1.369301,
+      "emos_mean_mae": 1.3725,
       "legacy_bucket_hit_rate": 0.75,
       "emos_bucket_hit_rate": 0.75
     },
     "paris": {
       "samples": 7,
       "legacy_mean_crps": 2.430082,
-      "emos_mean_crps": 2.365416,
+      "emos_mean_crps": 2.430082,
       "legacy_mean_mae": 2.518571,
-      "emos_mean_mae": 2.48437,
+      "emos_mean_mae": 2.518571,
       "legacy_bucket_hit_rate": 0.571429,
       "emos_bucket_hit_rate": 0.571429
     },
     "sao paulo": {
       "samples": 5,
       "legacy_mean_crps": 2.454756,
-      "emos_mean_crps": 2.410624,
+      "emos_mean_crps": 2.454756,
       "legacy_mean_mae": 2.628,
-      "emos_mean_mae": 2.689909,
+      "emos_mean_mae": 2.628,
       "legacy_bucket_hit_rate": 0.6,
       "emos_bucket_hit_rate": 0.6
     },
     "seattle": {
       "samples": 4,
       "legacy_mean_crps": 0.531656,
-      "emos_mean_crps": 0.457257,
+      "emos_mean_crps": 0.452784,
       "legacy_mean_mae": 0.0,
-      "emos_mean_mae": 0.11262,
+      "emos_mean_mae": 0.0,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 1.0
     },
     "seoul": {
       "samples": 6,
       "legacy_mean_crps": 0.328088,
-      "emos_mean_crps": 0.332722,
+      "emos_mean_crps": 0.328088,
       "legacy_mean_mae": 0.2,
-      "emos_mean_mae": 0.246659,
+      "emos_mean_mae": 0.2,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 1.0
     },
     "shanghai": {
       "samples": 2,
       "legacy_mean_crps": 0.250034,
-      "emos_mean_crps": 0.249952,
+      "emos_mean_crps": 0.250034,
       "legacy_mean_mae": 0.15,
-      "emos_mean_mae": 0.204362,
+      "emos_mean_mae": 0.15,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 1.0
     },
     "singapore": {
       "samples": 2,
       "legacy_mean_crps": 0.281993,
-      "emos_mean_crps": 0.270875,
+      "emos_mean_crps": 0.281993,
       "legacy_mean_mae": 0.15,
-      "emos_mean_mae": 0.070547,
+      "emos_mean_mae": 0.15,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 1.0
     },
     "taipei": {
       "samples": 3,
       "legacy_mean_crps": 0.356996,
-      "emos_mean_crps": 0.351432,
+      "emos_mean_crps": 0.356996,
       "legacy_mean_mae": 0.1,
-      "emos_mean_mae": 0.084784,
+      "emos_mean_mae": 0.1,
       "legacy_bucket_hit_rate": 1.0,
-      "emos_bucket_hit_rate": 0.666667
+      "emos_bucket_hit_rate": 1.0
     },
     "tel aviv": {
       "samples": 2,
       "legacy_mean_crps": 0.446758,
-      "emos_mean_crps": 0.434936,
+      "emos_mean_crps": 0.446758,
       "legacy_mean_mae": 0.3,
-      "emos_mean_mae": 0.196589,
+      "emos_mean_mae": 0.3,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 1.0
     },
     "tokyo": {
       "samples": 2,
       "legacy_mean_crps": 0.450128,
-      "emos_mean_crps": 0.429389,
+      "emos_mean_crps": 0.450128,
       "legacy_mean_mae": 0.25,
-      "emos_mean_mae": 0.100782,
+      "emos_mean_mae": 0.25,
       "legacy_bucket_hit_rate": 0.5,
       "emos_bucket_hit_rate": 1.0
     },
     "toronto": {
       "samples": 5,
       "legacy_mean_crps": 2.647861,
-      "emos_mean_crps": 2.512236,
+      "emos_mean_crps": 2.566068,
       "legacy_mean_mae": 2.532,
-      "emos_mean_mae": 2.535078,
+      "emos_mean_mae": 2.532,
       "legacy_bucket_hit_rate": 0.6,
       "emos_bucket_hit_rate": 0.6
     },
     "warsaw": {
       "samples": 3,
       "legacy_mean_crps": 1.618875,
-      "emos_mean_crps": 1.545129,
+      "emos_mean_crps": 1.618875,
       "legacy_mean_mae": 2.056667,
-      "emos_mean_mae": 1.942917,
+      "emos_mean_mae": 2.056667,
       "legacy_bucket_hit_rate": 0.333333,
       "emos_bucket_hit_rate": 0.333333
     },
     "wellington": {
       "samples": 6,
       "legacy_mean_crps": 0.266349,
-      "emos_mean_crps": 0.264506,
+      "emos_mean_crps": 0.266349,
       "legacy_mean_mae": 0.2,
-      "emos_mean_mae": 0.205623,
+      "emos_mean_mae": 0.2,
       "legacy_bucket_hit_rate": 1.0,
       "emos_bucket_hit_rate": 1.0
     }
diff --git a/artifacts/probability_calibration/training_samples.json b/artifacts/probability_calibration/training_samples.json
index 721c2ddb..f9991ffa 100644
--- a/artifacts/probability_calibration/training_samples.json
+++ b/artifacts/probability_calibration/training_samples.json
@@ -1,5 +1,7 @@
 {
   "sample_count": 105,
+  "snapshot_sample_count": 0,
+  "daily_record_sample_count": 105,
   "filled_actual_from_history": 2,
   "samples": [
     {
@@ -12,7 +14,8 @@
       "ens_median": 5.9,
       "ensemble_spread": 1.5499999999999998,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "ankara",
@@ -24,7 +27,8 @@
       "ens_median": 7.6,
       "ensemble_spread": 1.75,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "ankara",
@@ -36,7 +40,8 @@
       "ens_median": 7.9,
       "ensemble_spread": 1.5499999999999998,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "ankara",
@@ -48,7 +53,8 @@
       "ens_median": 10.0,
       "ensemble_spread": 0.8500000000000005,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "ankara",
@@ -60,7 +66,8 @@
       "ens_median": 13.9,
       "ensemble_spread": 1.75,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "ankara",
@@ -72,7 +79,8 @@
       "ens_median": 15.8,
       "ensemble_spread": 1.2000000000000002,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "ankara",
@@ -84,7 +92,8 @@
       "ens_median": 10.1,
       "ensemble_spread": 2.05,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "london",
@@ -96,7 +105,8 @@
       "ens_median": 10.1,
       "ensemble_spread": 0.9000000000000004,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "london",
@@ -108,7 +118,8 @@
       "ens_median": 11.4,
       "ensemble_spread": 2.1500000000000004,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "london",
@@ -120,7 +131,8 @@
       "ens_median": 15.4,
       "ensemble_spread": 0.6,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "london",
@@ -132,7 +144,8 @@
       "ens_median": 13.3,
       "ensemble_spread": 0.9500000000000002,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "london",
@@ -144,7 +157,8 @@
       "ens_median": 16.8,
       "ensemble_spread": 1.299999999999999,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "london",
@@ -156,7 +170,8 @@
       "ens_median": 15.8,
       "ensemble_spread": 0.6,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "new york",
@@ -168,7 +183,8 @@
       "ens_median": 46.7,
       "ensemble_spread": 9.7,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "new york",
@@ -180,7 +196,8 @@
       "ens_median": 67.2,
       "ensemble_spread": 2.1499999999999986,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "new york",
@@ -192,7 +209,8 @@
       "ens_median": 66.6,
       "ensemble_spread": 4.149999999999999,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "new york",
@@ -204,7 +222,8 @@
       "ens_median": 37.0,
       "ensemble_spread": 2.25,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "paris",
@@ -216,7 +235,8 @@
       "ens_median": 16.7,
       "ensemble_spread": 0.7000000000000011,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "paris",
@@ -228,7 +248,8 @@
       "ens_median": 16.2,
       "ensemble_spread": 1.5999999999999996,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "paris",
@@ -240,7 +261,8 @@
       "ens_median": 17.4,
       "ensemble_spread": 0.6999999999999993,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "paris",
@@ -252,7 +274,8 @@
       "ens_median": 15.3,
       "ensemble_spread": 0.6,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "paris",
@@ -264,7 +287,8 @@
       "ens_median": 13.1,
       "ensemble_spread": 0.8499999999999996,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "paris",
@@ -276,7 +300,8 @@
       "ens_median": 14.8,
       "ensemble_spread": 0.9000000000000004,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "paris",
@@ -288,7 +313,8 @@
       "ens_median": 16.3,
       "ensemble_spread": 0.6,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "seoul",
@@ -300,7 +326,8 @@
       "ens_median": 4.0,
       "ensemble_spread": 0.7999999999999998,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "seoul",
@@ -312,7 +339,8 @@
       "ens_median": 6.3,
       "ensemble_spread": 1.4,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "seoul",
@@ -324,7 +352,8 @@
       "ens_median": 4.9,
       "ensemble_spread": 1.35,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "seoul",
@@ -336,7 +365,8 @@
       "ens_median": 9.1,
       "ensemble_spread": 1.7999999999999998,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "seoul",
@@ -348,7 +378,8 @@
       "ens_median": 7.4,
       "ensemble_spread": 0.7999999999999998,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "seoul",
@@ -360,7 +391,8 @@
       "ens_median": 6.3,
       "ensemble_spread": 1.7999999999999998,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "toronto",
@@ -372,7 +404,8 @@
       "ens_median": 16.9,
       "ensemble_spread": 2.8500000000000005,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "toronto",
@@ -384,7 +417,8 @@
       "ens_median": 12.3,
       "ensemble_spread": 2.0,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "toronto",
@@ -396,7 +430,8 @@
       "ens_median": 17.6,
       "ensemble_spread": 4.75,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "toronto",
@@ -408,7 +443,8 @@
       "ens_median": -1.1,
       "ensemble_spread": 0.8,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "toronto",
@@ -420,7 +456,8 @@
       "ens_median": 6.3,
       "ensemble_spread": 2.1500000000000004,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "buenos aires",
@@ -432,7 +469,8 @@
       "ens_median": 26.0,
       "ensemble_spread": 1.3999999999999986,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "buenos aires",
@@ -444,7 +482,8 @@
       "ens_median": 26.4,
       "ensemble_spread": 1.200000000000001,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "buenos aires",
@@ -456,7 +495,8 @@
       "ens_median": 26.2,
       "ensemble_spread": 1.5999999999999996,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "buenos aires",
@@ -468,7 +508,8 @@
       "ens_median": 26.0,
       "ensemble_spread": 2.0,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "buenos aires",
@@ -480,7 +521,8 @@
       "ens_median": 27.9,
       "ensemble_spread": 2.0999999999999996,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "wellington",
@@ -492,7 +534,8 @@
       "ens_median": 18.4,
       "ensemble_spread": 0.6,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "wellington",
@@ -504,7 +547,8 @@
       "ens_median": 17.2,
       "ensemble_spread": 0.6,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "wellington",
@@ -516,7 +560,8 @@
       "ens_median": 15.7,
       "ensemble_spread": 0.9000000000000004,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "wellington",
@@ -528,7 +573,8 @@
       "ens_median": 18.7,
       "ensemble_spread": 1.049999999999999,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "wellington",
@@ -540,7 +586,8 @@
       "ens_median": 19.1,
       "ensemble_spread": 0.9500000000000011,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "wellington",
@@ -552,7 +599,8 @@
       "ens_median": 17.1,
       "ensemble_spread": 2.0999999999999996,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "chicago",
@@ -564,7 +612,8 @@
       "ens_median": 62.8,
       "ensemble_spread": 1.3000000000000007,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "chicago",
@@ -576,7 +625,8 @@
       "ens_median": 60.1,
       "ensemble_spread": 5.600000000000001,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "chicago",
@@ -588,7 +638,8 @@
       "ens_median": 70.8,
       "ensemble_spread": 10.799999999999997,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "chicago",
@@ -600,7 +651,8 @@
       "ens_median": 44.2,
       "ensemble_spread": 5.350000000000001,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "sao paulo",
@@ -612,7 +664,8 @@
       "ens_median": 27.4,
       "ensemble_spread": 1.5,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "sao paulo",
@@ -624,7 +677,8 @@
       "ens_median": 28.0,
       "ensemble_spread": 2.1999999999999993,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "sao paulo",
@@ -636,7 +690,8 @@
       "ens_median": 23.5,
       "ensemble_spread": 1.4000000000000004,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "sao paulo",
@@ -648,7 +703,8 @@
       "ens_median": 29.4,
       "ensemble_spread": 2.450000000000001,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "sao paulo",
@@ -660,7 +716,8 @@
       "ens_median": 26.5,
       "ensemble_spread": 1.200000000000001,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "dallas",
@@ -672,7 +729,8 @@
       "ens_median": 72.7,
       "ensemble_spread": 2.1499999999999986,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "dallas",
@@ -684,7 +742,8 @@
       "ens_median": 71.5,
       "ensemble_spread": 6.200000000000003,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "dallas",
@@ -696,7 +755,8 @@
       "ens_median": 81.0,
       "ensemble_spread": 3.8500000000000014,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "dallas",
@@ -708,7 +768,8 @@
       "ens_median": 74.4,
       "ensemble_spread": 9.299999999999997,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "miami",
@@ -720,7 +781,8 @@
       "ens_median": 81.6,
       "ensemble_spread": 1.3999999999999986,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "miami",
@@ -732,7 +794,8 @@
       "ens_median": 82.7,
       "ensemble_spread": 1.5500000000000043,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "miami",
@@ -744,7 +807,8 @@
       "ens_median": 83.9,
       "ensemble_spread": 1.5499999999999972,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "miami",
@@ -756,7 +820,8 @@
       "ens_median": 70.4,
       "ensemble_spread": 2.8999999999999986,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "miami",
@@ -768,7 +833,8 @@
       "ens_median": 74.0,
       "ensemble_spread": 2.5500000000000043,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "atlanta",
@@ -780,7 +846,8 @@
       "ens_median": 79.7,
       "ensemble_spread": 4.600000000000001,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "atlanta",
@@ -792,7 +859,8 @@
       "ens_median": 67.7,
       "ensemble_spread": 5.899999999999999,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "atlanta",
@@ -804,7 +872,8 @@
       "ens_median": 77.4,
       "ensemble_spread": 2.6000000000000014,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "atlanta",
@@ -816,7 +885,8 @@
       "ens_median": 52.2,
       "ensemble_spread": 4.0,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "atlanta",
@@ -828,7 +898,8 @@
       "ens_median": 65.9,
       "ensemble_spread": 1.5500000000000007,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "seattle",
@@ -840,7 +911,8 @@
       "ens_median": 52.8,
       "ensemble_spread": 1.3999999999999986,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "seattle",
@@ -852,7 +924,8 @@
       "ens_median": 50.9,
       "ensemble_spread": 4.350000000000001,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "seattle",
@@ -864,7 +937,8 @@
       "ens_median": 45.1,
       "ensemble_spread": 2.0,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "seattle",
@@ -876,7 +950,8 @@
       "ens_median": 55.9,
       "ensemble_spread": 1.3500000000000014,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "lucknow",
@@ -888,7 +963,8 @@
       "ens_median": 33.9,
       "ensemble_spread": 1.8500000000000014,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "lucknow",
@@ -900,7 +976,8 @@
       "ens_median": 34.6,
       "ensemble_spread": 2.0,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "lucknow",
@@ -912,7 +989,8 @@
       "ens_median": 34.6,
       "ensemble_spread": 1.3000000000000007,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "lucknow",
@@ -924,7 +1002,8 @@
       "ens_median": 34.7,
       "ensemble_spread": 1.75,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "munich",
@@ -936,7 +1015,8 @@
       "ens_median": 15.1,
       "ensemble_spread": 1.6000000000000005,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "munich",
@@ -948,7 +1028,8 @@
       "ens_median": 14.3,
       "ensemble_spread": 0.9000000000000004,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "munich",
@@ -960,7 +1041,8 @@
       "ens_median": 14.5,
       "ensemble_spread": 0.8999999999999995,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "munich",
@@ -972,7 +1054,8 @@
       "ens_median": 15.7,
       "ensemble_spread": 1.2999999999999998,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "munich",
@@ -984,7 +1067,8 @@
       "ens_median": 10.6,
       "ensemble_spread": 0.6,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "munich",
@@ -996,7 +1080,8 @@
       "ens_median": 11.8,
       "ensemble_spread": 1.3000000000000007,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "hong kong",
@@ -1008,7 +1093,8 @@
       "ens_median": 23.1,
       "ensemble_spread": 1.9000000000000004,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "hong kong",
@@ -1020,7 +1106,8 @@
       "ens_median": 24.9,
       "ensemble_spread": 0.6,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "hong kong",
@@ -1032,7 +1119,8 @@
       "ens_median": 25.0,
       "ensemble_spread": 0.6,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "taipei",
@@ -1044,7 +1132,8 @@
       "ens_median": 25.4,
       "ensemble_spread": 2.25,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "taipei",
@@ -1056,7 +1145,8 @@
       "ens_median": 27.5,
       "ensemble_spread": 1.0500000000000007,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "taipei",
@@ -1068,7 +1158,8 @@
       "ens_median": 21.3,
       "ensemble_spread": 1.1500000000000004,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "milan",
@@ -1080,7 +1171,8 @@
       "ens_median": 15.1,
       "ensemble_spread": 9.1,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "milan",
@@ -1092,7 +1184,8 @@
       "ens_median": 13.8,
       "ensemble_spread": 0.6,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "milan",
@@ -1104,7 +1197,8 @@
       "ens_median": 17.0,
       "ensemble_spread": 1.25,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "warsaw",
@@ -1116,7 +1210,8 @@
       "ens_median": 10.5,
       "ensemble_spread": 0.6499999999999995,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "warsaw",
@@ -1128,7 +1223,8 @@
       "ens_median": 14.2,
       "ensemble_spread": 1.4000000000000004,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "warsaw",
@@ -1140,7 +1236,8 @@
       "ens_median": 12.3,
       "ensemble_spread": 1.5999999999999996,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "singapore",
@@ -1152,7 +1249,8 @@
       "ens_median": 30.1,
       "ensemble_spread": 0.9500000000000011,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "singapore",
@@ -1164,7 +1262,8 @@
       "ens_median": 32.1,
       "ensemble_spread": 1.3499999999999996,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "shanghai",
@@ -1176,7 +1275,8 @@
       "ens_median": 13.0,
       "ensemble_spread": 1.1500000000000004,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "shanghai",
@@ -1188,7 +1288,8 @@
       "ens_median": 11.2,
       "ensemble_spread": 0.7999999999999998,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "tokyo",
@@ -1200,7 +1301,8 @@
       "ens_median": 15.8,
       "ensemble_spread": 1.5499999999999998,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "tokyo",
@@ -1212,7 +1314,8 @@
       "ens_median": 18.5,
       "ensemble_spread": 2.0999999999999996,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "tel aviv",
@@ -1224,7 +1327,8 @@
       "ens_median": 28.9,
       "ensemble_spread": 2.1500000000000004,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "tel aviv",
@@ -1236,7 +1340,8 @@
       "ens_median": 21.1,
       "ensemble_spread": 1.5,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "madrid",
@@ -1248,7 +1353,8 @@
       "ens_median": 18.7,
       "ensemble_spread": 1.8499999999999996,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     },
     {
       "city": "madrid",
@@ -1260,7 +1366,8 @@
       "ens_median": 18.8,
       "ensemble_spread": 1.8999999999999995,
       "max_so_far_gap": null,
-      "peak_flag": 0.0
+      "peak_flag": 0.0,
+      "sample_source": "daily_record"
     }
   ]
 }
\ No newline at end of file
diff --git a/data/probability_training_snapshots.jsonl b/data/probability_training_snapshots.jsonl
new file mode 100644
index 00000000..bb34e8e7
--- /dev/null
+++ b/data/probability_training_snapshots.jsonl
@@ -0,0 +1 @@
+{"city": "ankara", "timestamp": "2026-03-20T12:00:00+03:00", "date": "2026-03-20", "temp_symbol": "°C", "raw_mu": 15.2, "raw_sigma": 1.2, "deb_prediction": 15.4, "ensemble": {"p10": 14.8, "median": 15.8, "p90": 17.9}, "multi_model": {"ECMWF": 15.8, "GFS": 14.1, "ICON": 15.9}, "max_so_far": 15.0, "peak_status": "before", "prob_snapshot": [{"v": 15, "p": 0.552}, {"v": 16, "p": 0.377}], "shadow_prob_snapshot": [{"v": 15, "p": 0.324}, {"v": 16, "p": 0.238}], "probability_engine": "legacy", "probability_mode": "emos_shadow", "calibration_version": "emos-20260320130245", "calibration_source": "artifacts/probability_calibration/default.json", "calibrated_mu": 15.1, "calibrated_sigma": 1.25}
diff --git a/docs/PROBABILITY_SNAPSHOT_ARCHIVE_ZH.md b/docs/PROBABILITY_SNAPSHOT_ARCHIVE_ZH.md
new file mode 100644
index 00000000..00471ed5
--- /dev/null
+++ b/docs/PROBABILITY_SNAPSHOT_ARCHIVE_ZH.md
@@ -0,0 +1,324 @@
+# 概率训练样本归档说明（中文）
+
+## 1. 目的
+
+这份文档说明两件事：
+
+1. 为什么 `EMOS` 训练不能只依赖历史实测天气
+2. 未来如何持续沉淀“历史预测记录”，让概率引擎越训越稳
+
+一句话结论：
+
+- 历史实测天气只能补 `actual_high`
+- 真正决定 `EMOS` 训练质量的是“当时那一刻的预测快照”
+
+## 2. 什么是“历史预测记录”
+
+对 PolyWeather 来说，一条可训练的历史预测记录，至少应该包含这些字段：
+
+- `city`
+- `timestamp`
+- `date`
+- `raw_mu`
+- `raw_sigma`
+- `deb_prediction`
+- `ensemble p10 / p50 / p90`
+- `multi-model forecasts`
+- `max_so_far`
+- `peak_status`
+- `prob_snapshot`
+- 当天最终 `actual_high`
+- 当天最终 `settlement bucket`
+
+这类记录的核心价值是：
+
+- 还原“当时系统实际看到什么”
+- 再对照“后来真实发生了什么”
+
+只有这两者成对，`EMOS` 才能学习偏差。
+
+## 3. 为什么不能只用历史天气实测
+
+历史天气 CSV 只能告诉你：
+
+- 当天最高温是多少
+- 某小时温度是多少
+
+但它不能告诉你：
+
+- 当天早上 09:00 时，系统的 `mu` 是多少
+- 当时的 `ensemble spread` 是多少
+- 当时 `DEB` 怎么看
+- 当时的 top bucket 是什么
+
+所以：
+
+- 历史实测天气是标签
+- 历史预测记录才是训练输入
+
+缺少后者，EMOS 只能学到很有限的东西。
+
+## 4. 当前项目里已经有的基础
+
+### 4.1 已有历史日记录
+
+文件：
+
+- [daily_records.json](/E:/web/PolyWeather/data/daily_records.json)
+
+当前已经保存了一部分训练相关字段，例如：
+
+- `forecasts`
+- `actual_high`
+- `deb_prediction`
+- `mu`
+- `prob_snapshot`
+- `shadow_prob_snapshot`
+- `probability_calibration`
+- `probability_features`
+
+这已经是“历史预测记录”的雏形。
+
+### 4.2 已有历史天气 CSV
+
+目录：
+
+- [data/historical](/E:/web/PolyWeather/data/historical)
+
+它们可以帮助补：
+
+- `actual_high`
+- `settlement history`
+
+但不能替代预测快照归档。
+
+## 5. 未来应该怎么存历史预测记录
+
+推荐做法是：
+
+### 5.1 固定时点归档
+
+每天为每个重点城市固定存几次快照，例如：
+
+- 当地 `09:00`
+- 当地 `12:00`
+- 当地 `15:00`
+
+这样能确保每个交易日都有稳定可比样本。
+
+### 5.2 关键变化时补充归档
+
+除了固定时点，还应该在以下情况额外存一次：
+
+- `max_so_far` 创新高
+- `mu` 变化超过阈值
+- `top bucket` 发生变化
+- `shadow top bucket` 发生变化
+
+这样能捕捉真正有训练价值的转折点。
+
+### 5.3 建议的存储格式
+
+建议新增一个文件，例如：
+
+- `data/probability_training_snapshots.jsonl`
+
+每一行保存一条 JSON 记录。
+
+优点：
+
+- 追加写入简单
+- 后续导出训练集方便
+- 不容易因为单个大 JSON 文件损坏而全盘受影响
+
+## 6. 一条建议的快照结构
+
+示例：
+
+```json
+{
+  "city": "ankara",
+  "timestamp": "2026-03-20T12:00:00+03:00",
+  "date": "2026-03-20",
+  "raw_mu": 15.2,
+  "raw_sigma": 1.2,
+  "deb_prediction": 15.4,
+  "ensemble": {
+    "p10": 14.8,
+    "median": 15.8,
+    "p90": 17.9
+  },
+  "multi_model": {
+    "ECMWF": 15.8,
+    "GFS": 14.1,
+    "ICON": 15.9,
+    "GEM": 16.5,
+    "JMA": 14.5
+  },
+  "max_so_far": 15.0,
+  "peak_status": "before",
+  "prob_snapshot": [
+    {"v": 15, "p": 0.552},
+    {"v": 16, "p": 0.377}
+  ],
+  "shadow_prob_snapshot": [
+    {"v": 15, "p": 0.324},
+    {"v": 16, "p": 0.238}
+  ],
+  "probability_engine": "legacy",
+  "probability_mode": "emos_shadow",
+  "calibration_version": "emos-20260320130245"
+}
+```
+
+当天结束后，再由后处理脚本回填：
+
+- `actual_high`
+- `settlement_bucket`
+
+## 7. 现阶段你可以执行的命令
+
+### 7.1 回填历史天气 CSV
+
+```bash
+python scripts/backfill_historical_weather.py
+```
+
+作用：
+
+- 补全 30 城市历史天气时序 CSV
+
+### 7.2 从历史 CSV 构建日级结算标签
+
+```bash
+python scripts/build_settlement_history_from_csv.py
+```
+
+作用：
+
+- 生成 [settlement_history.json](/E:/web/PolyWeather/artifacts/probability_calibration/settlement_history.json)
+
+### 7.3 导出当前训练样本
+
+```bash
+python scripts/export_probability_training_dataset.py
+```
+
+作用：
+
+- 生成 [training_samples.json](/E:/web/PolyWeather/artifacts/probability_calibration/training_samples.json)
+
+### 7.4 重训 EMOS
+
+```bash
+python scripts/fit_probability_calibration.py
+```
+
+作用：
+
+- 生成新的 [default.json](/E:/web/PolyWeather/artifacts/probability_calibration/default.json)
+
+### 7.5 离线评估训练效果
+
+```bash
+python scripts/evaluate_probability_calibration.py
+```
+
+作用：
+
+- 生成 [evaluation_report.json](/E:/web/PolyWeather/artifacts/probability_calibration/evaluation_report.json)
+
+### 7.6 回填 shadow 结果到历史记录
+
+```bash
+python scripts/backfill_probability_shadow_history.py
+```
+
+作用：
+
+- 把 `shadow_prob_snapshot` 和 `probability_calibration` 回填到 [daily_records.json](/E:/web/PolyWeather/data/daily_records.json)
+
+### 7.7 生成线上 shadow 滚动报表
+
+```bash
+python scripts/build_probability_shadow_report.py
+```
+
+作用：
+
+- 生成 [shadow_report.json](/E:/web/PolyWeather/artifacts/probability_calibration/shadow_report.json)
+
+## 8. 推荐的一整套重训流程
+
+如果过了十天、半个月，想重新训练一次，建议按这个顺序执行：
+
+```bash
+python scripts/build_settlement_history_from_csv.py
+python scripts/export_probability_training_dataset.py
+python scripts/fit_probability_calibration.py
+python scripts/evaluate_probability_calibration.py
+python scripts/backfill_probability_shadow_history.py
+python scripts/build_probability_shadow_report.py
+```
+
+如果历史天气 CSV 还没补全，再先执行：
+
+```bash
+python scripts/backfill_historical_weather.py
+```
+
+## 9. 怎么判断这次训练有没有进步
+
+重训后，不要只看一个指标。
+
+至少看这 4 个：
+
+1. `CRPS`
+- 越低越好
+
+2. `MAE`
+- 越低越好
+- 至少不要明显变差
+
+3. `Bucket Hit Rate`
+- 越高越好
+- 这是业务上非常关键的指标
+
+4. `Bucket Brier`
+- 越低越好
+- 反映概率分布质量
+
+只有同时满足下面条件，才可以说训练效果真的进步：
+
+- `CRPS` 下降
+- `MAE` 不上升
+- `Bucket Hit Rate` 不下降
+- `Bucket Brier` 不上升
+
+## 10. 当前最重要的现实判断
+
+过去的“完整历史预测记录”通常没法完全补出来，除非：
+
+1. 你之前就存过
+2. 你接入了支持 forecast archive 的商业数据源
+
+所以现实里最重要的不是“把过去全补齐”，而是：
+
+- 从现在开始系统化归档
+- 每天稳定沉淀可训练样本
+- 定期离线重训
+
+## 11. 推荐的下一步
+
+最值得做的改造是：
+
+1. 新增 `probability_training_snapshots.jsonl`
+2. 每次分析时自动追加一条快照
+3. 当天结束后自动回填 `actual_high`
+4. 每 1-2 周重新训练一次
+
+## 12. 总结
+
+如果只记住一句话，就记这个：
+
+**EMOS 要想越训越好，关键不是多下载一点历史天气，而是持续保存“当时系统看到的预测快照”。**
diff --git a/scripts/export_probability_training_dataset.py b/scripts/export_probability_training_dataset.py
index 420f8785..69890a5a 100644
--- a/scripts/export_probability_training_dataset.py
+++ b/scripts/export_probability_training_dataset.py
@@ -7,10 +7,11 @@ PROJECT_ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
 if PROJECT_ROOT not in sys.path:
     sys.path.insert(0, PROJECT_ROOT)
 
-from src.analysis.deb_algorithm import load_history  # noqa: E402
 from scripts.fit_probability_calibration import (  # noqa: E402
     _extract_samples,
+    _load_history_with_fallback,
     _load_json_if_exists,
+    _load_snapshot_rows,
 )
 
 
@@ -38,20 +39,30 @@ def main():
             "training_samples.json",
         ),
     )
+    parser.add_argument(
+        "--snapshot-file",
+        default=os.path.join(PROJECT_ROOT, "data", "probability_training_snapshots.jsonl"),
+    )
     args = parser.parse_args()
 
-    history = load_history(args.history_file)
+    history = _load_history_with_fallback(args.history_file)
     settlement_history = _load_json_if_exists(args.settlement_history)
+    snapshot_rows = _load_snapshot_rows(args.snapshot_file)
     samples, filled_actual_from_history = _extract_samples(
         history,
         settlement_history=settlement_history,
+        snapshot_rows=snapshot_rows,
     )
+    snapshot_count = sum(1 for sample in samples if sample.get("sample_source") == "snapshot")
+    daily_record_count = sum(1 for sample in samples if sample.get("sample_source") == "daily_record")
 
     output_dir = os.path.dirname(os.path.abspath(args.output))
     if output_dir:
         os.makedirs(output_dir, exist_ok=True)
     payload = {
         "sample_count": len(samples),
+        "snapshot_sample_count": snapshot_count,
+        "daily_record_sample_count": daily_record_count,
         "filled_actual_from_history": filled_actual_from_history,
         "samples": samples,
     }
diff --git a/scripts/fit_probability_calibration.py b/scripts/fit_probability_calibration.py
index b67095f5..7a27a4c3 100644
--- a/scripts/fit_probability_calibration.py
+++ b/scripts/fit_probability_calibration.py
@@ -33,11 +33,128 @@ def _load_json_if_exists(path):
     return data if isinstance(data, dict) else {}
 
 
-def _extract_samples(history, settlement_history=None):
+def _load_history_with_fallback(path):
+    data = load_history(path)
+    if data:
+        return data
+    return _load_json_if_exists(path)
+
+
+def _load_snapshot_rows(path):
+    rows = []
+    if not path or not os.path.exists(path):
+        return rows
+    with open(path, "r", encoding="utf-8") as fh:
+        for line in fh:
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                row = json.loads(line)
+            except Exception:
+                continue
+            if isinstance(row, dict):
+                rows.append(row)
+    return rows
+
+
+def _actual_high_for(history, settlement_history, city, date_str):
+    city_rows = (history or {}).get(city) or {}
+    record = city_rows.get(date_str) or {}
+    actual_high = _sf(record.get("actual_high")) if isinstance(record, dict) else None
+    filled = False
+    if actual_high is None:
+        actual_high = _sf(((settlement_history.get(city) or {}).get(date_str) or {}).get("max_temp"))
+        filled = actual_high is not None
+    return actual_high, filled
+
+
+def _extract_snapshot_samples(history, snapshot_rows, settlement_history=None):
     samples = []
     filled_actual_from_history = 0
     today = datetime.utcnow().strftime("%Y-%m-%d")
     settlement_history = settlement_history or {}
+
+    for row in snapshot_rows or []:
+        city = str(row.get("city") or "").strip().lower()
+        date_str = str(row.get("date") or "").strip()
+        if not city or not date_str or date_str == today:
+            continue
+
+        actual_high, filled = _actual_high_for(history, settlement_history, city, date_str)
+        if actual_high is None:
+            continue
+        if filled:
+            filled_actual_from_history += 1
+
+        raw_mu = _sf(row.get("raw_mu"))
+        raw_sigma = _sf(row.get("raw_sigma"))
+        deb_prediction = _sf(row.get("deb_prediction"))
+        ensemble = row.get("ensemble") or {}
+        if not isinstance(ensemble, dict):
+            ensemble = {}
+        ens_median = _sf(ensemble.get("median"))
+        ensemble_spread = None
+        ens_p10 = _sf(ensemble.get("p10"))
+        ens_p90 = _sf(ensemble.get("p90"))
+        if ens_p10 is not None and ens_p90 is not None and ens_p90 >= ens_p10:
+            ensemble_spread = max(0.1, (ens_p90 - ens_p10) / 2.56)
+        multi_model = row.get("multi_model") or {}
+        if not isinstance(multi_model, dict):
+            multi_model = {}
+        forecast_values = [val for val in (_sf(v) for v in multi_model.values()) if val is not None]
+        forecast_values.sort()
+        if ensemble_spread is None:
+            if len(forecast_values) >= 2:
+                ensemble_spread = max(0.6, (forecast_values[-1] - forecast_values[0]) / 2.0)
+            elif raw_sigma is not None:
+                ensemble_spread = raw_sigma
+            else:
+                ensemble_spread = 1.0
+        if raw_sigma is None:
+            raw_sigma = ensemble_spread
+
+        peak_status = str(row.get("peak_status") or "before").strip().lower()
+        peak_flag = 0.0
+        if peak_status == "in_window":
+            peak_flag = 0.5
+        elif peak_status == "past":
+            peak_flag = 1.0
+
+        max_so_far = _sf(row.get("max_so_far"))
+        max_so_far_gap = None
+        if deb_prediction is not None and max_so_far is not None:
+            max_so_far_gap = deb_prediction - max_so_far
+
+        if raw_mu is None:
+            continue
+
+        samples.append(
+            {
+                "city": city,
+                "date": date_str,
+                "timestamp": row.get("timestamp"),
+                "actual_high": actual_high,
+                "raw_mu": raw_mu,
+                "raw_sigma": raw_sigma or 1.0,
+                "deb_prediction": deb_prediction,
+                "ens_median": ens_median if ens_median is not None else raw_mu,
+                "ensemble_spread": ensemble_spread,
+                "max_so_far_gap": max_so_far_gap,
+                "peak_flag": peak_flag,
+                "sample_source": "snapshot",
+            }
+        )
+
+    return samples, filled_actual_from_history
+
+
+def _extract_daily_record_samples(history, settlement_history=None, excluded_keys=None):
+    samples = []
+    filled_actual_from_history = 0
+    today = datetime.utcnow().strftime("%Y-%m-%d")
+    settlement_history = settlement_history or {}
+    excluded_keys = excluded_keys or set()
     for city, city_rows in (history or {}).items():
         if not isinstance(city_rows, dict):
             continue
@@ -45,6 +162,8 @@ def _extract_samples(history, settlement_history=None):
         for date_str, record in city_rows.items():
             if date_str == today or not isinstance(record, dict):
                 continue
+            if (city, date_str) in excluded_keys:
+                continue
             actual_high = _sf(record.get("actual_high"))
             if actual_high is None:
                 actual_high = _sf((city_settlement.get(date_str) or {}).get("max_temp"))
@@ -99,11 +218,30 @@ def _extract_samples(history, settlement_history=None):
                     "ensemble_spread": ensemble_spread,
                     "max_so_far_gap": max_so_far_gap,
                     "peak_flag": peak_flag,
+                    "sample_source": "daily_record",
                 }
             )
     return samples, filled_actual_from_history
 
 
+def _extract_samples(history, settlement_history=None, snapshot_rows=None):
+    snapshot_samples, snapshot_filled = _extract_snapshot_samples(
+        history,
+        snapshot_rows or [],
+        settlement_history=settlement_history,
+    )
+    excluded_keys = {
+        (sample["city"], sample["date"])
+        for sample in snapshot_samples
+    }
+    daily_samples, daily_filled = _extract_daily_record_samples(
+        history,
+        settlement_history=settlement_history,
+        excluded_keys=excluded_keys,
+    )
+    return snapshot_samples + daily_samples, snapshot_filled + daily_filled
+
+
 def main():
     parser = argparse.ArgumentParser(description="Fit PolyWeather probability calibration parameters.")
     parser.add_argument(
@@ -126,6 +264,11 @@ def main():
         ),
         help="Optional daily settlement history JSON built from historical CSV files.",
     )
+    parser.add_argument(
+        "--snapshot-file",
+        default=os.path.join(PROJECT_ROOT, "data", "probability_training_snapshots.jsonl"),
+        help="Optional JSONL file with archived probability snapshots.",
+    )
     parser.add_argument(
         "--version",
         default=None,
@@ -133,11 +276,13 @@ def main():
     )
     args = parser.parse_args()
 
-    history = load_history(args.history_file)
+    history = _load_history_with_fallback(args.history_file)
     settlement_history = _load_json_if_exists(args.settlement_history)
+    snapshot_rows = _load_snapshot_rows(args.snapshot_file)
     samples, filled_actual_from_history = _extract_samples(
         history,
         settlement_history=settlement_history,
+        snapshot_rows=snapshot_rows,
     )
     calibration = fit_calibration(samples, version=args.version)
     if not samples:
diff --git a/src/analysis/probability_snapshot_archive.py b/src/analysis/probability_snapshot_archive.py
new file mode 100644
index 00000000..c5964d32
--- /dev/null
+++ b/src/analysis/probability_snapshot_archive.py
@@ -0,0 +1,187 @@
+from __future__ import annotations
+
+import json
+import os
+from datetime import datetime
+from typing import Any, Dict, List, Optional
+
+DEDUP_SCAN_LINES = 200
+MU_THRESHOLD = 0.2
+SIGMA_THRESHOLD = 0.15
+MAX_SO_FAR_THRESHOLD = 0.2
+
+
+def _sf(value: Any) -> Optional[float]:
+    if value is None:
+        return None
+    try:
+        return float(value)
+    except Exception:
+        return None
+
+
+def _compact_snapshot(distribution: Optional[List[Dict[str, Any]]]) -> List[Dict[str, Any]]:
+    compact: List[Dict[str, Any]] = []
+    for row in distribution or []:
+        if not isinstance(row, dict):
+            continue
+        value = row.get("value")
+        probability = row.get("probability")
+        if value is None or probability is None:
+            continue
+        try:
+            compact.append(
+                {
+                    "v": int(value),
+                    "p": round(float(probability), 3),
+                }
+            )
+        except Exception:
+            continue
+        if len(compact) >= 4:
+            break
+    return compact
+
+
+def _top_bucket(snapshot: Optional[List[Dict[str, Any]]]) -> Optional[int]:
+    best_value = None
+    best_prob = -1.0
+    for row in snapshot or []:
+        if not isinstance(row, dict):
+            continue
+        value = row.get("v")
+        prob = _sf(row.get("p"))
+        if value is None or prob is None:
+            continue
+        if prob > best_prob:
+            best_value = int(value)
+            best_prob = prob
+    return best_value
+
+
+def _load_recent_rows(path: str, max_lines: int = DEDUP_SCAN_LINES) -> List[Dict[str, Any]]:
+    if not os.path.exists(path):
+        return []
+    with open(path, "r", encoding="utf-8") as fh:
+        lines = fh.readlines()[-max_lines:]
+    rows = []
+    for line in lines:
+        line = line.strip()
+        if not line:
+            continue
+        try:
+            row = json.loads(line)
+        except Exception:
+            continue
+        if isinstance(row, dict):
+            rows.append(row)
+    return rows
+
+
+def _should_skip_append(path: str, payload: Dict[str, Any]) -> bool:
+    recent_rows = _load_recent_rows(path)
+    city = payload.get("city")
+    date_str = payload.get("date")
+    if not city or not date_str:
+        return False
+
+    for row in reversed(recent_rows):
+        if row.get("city") != city or row.get("date") != date_str:
+            continue
+        if row.get("peak_status") != payload.get("peak_status"):
+            return False
+        if row.get("probability_mode") != payload.get("probability_mode"):
+            return False
+
+        current_top = _top_bucket(payload.get("prob_snapshot"))
+        previous_top = _top_bucket(row.get("prob_snapshot"))
+        current_shadow_top = _top_bucket(payload.get("shadow_prob_snapshot"))
+        previous_shadow_top = _top_bucket(row.get("shadow_prob_snapshot"))
+        if current_top != previous_top or current_shadow_top != previous_shadow_top:
+            return False
+
+        if abs((_sf(payload.get("raw_mu")) or 0.0) - (_sf(row.get("raw_mu")) or 0.0)) > MU_THRESHOLD:
+            return False
+        if abs((_sf(payload.get("raw_sigma")) or 0.0) - (_sf(row.get("raw_sigma")) or 0.0)) > SIGMA_THRESHOLD:
+            return False
+        if abs((_sf(payload.get("max_so_far")) or 0.0) - (_sf(row.get("max_so_far")) or 0.0)) > MAX_SO_FAR_THRESHOLD:
+            return False
+
+        return True
+
+    return False
+
+
+def append_probability_snapshot(
+    city_name: str,
+    *,
+    local_date: str,
+    observation_time: Optional[str],
+    temp_symbol: str,
+    raw_mu: Optional[float],
+    raw_sigma: Optional[float],
+    deb_prediction: Optional[float],
+    ens_data: Optional[Dict[str, Any]],
+    current_forecasts: Optional[Dict[str, Any]],
+    max_so_far: Optional[float],
+    peak_status: Optional[str],
+    probabilities: Optional[List[Dict[str, Any]]],
+    shadow_probabilities: Optional[List[Dict[str, Any]]],
+    calibration_summary: Optional[Dict[str, Any]],
+    archive_path: Optional[str] = None,
+) -> None:
+    city_key = str(city_name or "").strip().lower()
+    if not city_key:
+        return
+
+    root_dir = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+    path = archive_path or os.path.join(
+        root_dir,
+        "data",
+        "probability_training_snapshots.jsonl",
+    )
+
+    calibration_summary = calibration_summary or {}
+    ens_data = ens_data or {}
+    current_forecasts = current_forecasts or {}
+    timestamp = str(observation_time or datetime.utcnow().isoformat() + "Z").strip()
+
+    payload = {
+        "city": city_key,
+        "timestamp": timestamp,
+        "date": local_date,
+        "temp_symbol": temp_symbol,
+        "raw_mu": _sf(raw_mu),
+        "raw_sigma": _sf(raw_sigma),
+        "deb_prediction": _sf(deb_prediction),
+        "ensemble": {
+            "p10": _sf(ens_data.get("p10")),
+            "median": _sf(ens_data.get("median")),
+            "p90": _sf(ens_data.get("p90")),
+        },
+        "multi_model": {
+            key: _sf(value)
+            for key, value in current_forecasts.items()
+            if _sf(value) is not None
+        },
+        "max_so_far": _sf(max_so_far),
+        "peak_status": peak_status,
+        "prob_snapshot": _compact_snapshot(probabilities),
+        "shadow_prob_snapshot": _compact_snapshot(shadow_probabilities),
+        "probability_engine": calibration_summary.get("engine"),
+        "probability_mode": calibration_summary.get("mode"),
+        "calibration_version": calibration_summary.get("calibration_version"),
+        "calibration_source": calibration_summary.get("calibration_source"),
+        "calibrated_mu": _sf(calibration_summary.get("calibrated_mu")),
+        "calibrated_sigma": _sf(calibration_summary.get("calibrated_sigma")),
+    }
+
+    parent = os.path.dirname(os.path.abspath(path))
+    if parent:
+        os.makedirs(parent, exist_ok=True)
+
+    if _should_skip_append(path, payload):
+        return
+
+    with open(path, "a", encoding="utf-8") as fh:
+        fh.write(json.dumps(payload, ensure_ascii=False) + "\n")
diff --git a/src/analysis/trend_engine.py b/src/analysis/trend_engine.py
index 7adf8a28..10c47b31 100644
--- a/src/analysis/trend_engine.py
+++ b/src/analysis/trend_engine.py
@@ -19,6 +19,7 @@ from src.analysis.probability_calibration import (
     apply_probability_calibration,
     build_probability_features,
 )
+from src.analysis.probability_snapshot_archive import append_probability_snapshot
 from src.analysis.settlement_rounding import apply_city_settlement, is_exact_settlement_city
 from src.data_collection.city_registry import CITY_REGISTRY
 from src.data_collection.city_risk_profiles import get_city_risk_profile
@@ -733,6 +734,26 @@ def analyze_weather_trend(
     except Exception:
         pass
 
+    try:
+        append_probability_snapshot(
+            city_name=city_name or "",
+            local_date=local_date_str,
+            observation_time=obs_time_raw or local_time_full or None,
+            temp_symbol=temp_symbol,
+            raw_mu=calibration_summary.get("raw_mu"),
+            raw_sigma=calibration_summary.get("raw_sigma"),
+            deb_prediction=_deb_to_save,
+            ens_data=ens_data,
+            current_forecasts=current_forecasts,
+            max_so_far=max_so_far,
+            peak_status=peak_status,
+            probabilities=_prob_list,
+            shadow_probabilities=_shadow_prob_list,
+            calibration_summary=calibration_summary,
+        )
+    except Exception:
+        pass
+
     # === Build recent list for trend_info ===
     recent_list = []
     for tm, t in recent_temps[:4]:
diff --git a/tests/test_probability_snapshot_archive.py b/tests/test_probability_snapshot_archive.py
new file mode 100644
index 00000000..9be1d041
--- /dev/null
+++ b/tests/test_probability_snapshot_archive.py
@@ -0,0 +1,117 @@
+import json
+from pathlib import Path
+
+from src.analysis.probability_snapshot_archive import append_probability_snapshot
+
+
+def test_append_probability_snapshot_writes_jsonl(tmp_path: Path):
+    archive_path = tmp_path / "probability_training_snapshots.jsonl"
+
+    append_probability_snapshot(
+        city_name="ankara",
+        local_date="2026-03-20",
+        observation_time="2026-03-20T12:00:00+03:00",
+        temp_symbol="°C",
+        raw_mu=15.2,
+        raw_sigma=1.2,
+        deb_prediction=15.4,
+        ens_data={"p10": 14.8, "median": 15.8, "p90": 17.9},
+        current_forecasts={"ECMWF": 15.8, "GFS": 14.1},
+        max_so_far=15.0,
+        peak_status="before",
+        probabilities=[{"value": 15, "probability": 0.552}],
+        shadow_probabilities=[{"value": 15, "probability": 0.324}],
+        calibration_summary={
+            "engine": "legacy",
+            "mode": "emos_shadow",
+            "calibration_version": "emos-test",
+            "calibration_source": "artifacts/probability_calibration/default.json",
+            "calibrated_mu": 15.1,
+            "calibrated_sigma": 1.25,
+        },
+        archive_path=str(archive_path),
+    )
+
+    lines = archive_path.read_text(encoding="utf-8").strip().splitlines()
+    assert len(lines) == 1
+    payload = json.loads(lines[0])
+    assert payload["city"] == "ankara"
+    assert payload["date"] == "2026-03-20"
+    assert payload["raw_mu"] == 15.2
+    assert payload["ensemble"]["median"] == 15.8
+    assert payload["prob_snapshot"][0]["v"] == 15
+    assert payload["shadow_prob_snapshot"][0]["v"] == 15
+    assert payload["calibration_version"] == "emos-test"
+
+
+def test_append_probability_snapshot_skips_near_duplicate(tmp_path: Path):
+    archive_path = tmp_path / "probability_training_snapshots.jsonl"
+    kwargs = dict(
+        city_name="ankara",
+        local_date="2026-03-20",
+        observation_time="2026-03-20T12:00:00+03:00",
+        temp_symbol="°C",
+        raw_mu=15.2,
+        raw_sigma=1.2,
+        deb_prediction=15.4,
+        ens_data={"p10": 14.8, "median": 15.8, "p90": 17.9},
+        current_forecasts={"ECMWF": 15.8, "GFS": 14.1},
+        max_so_far=15.0,
+        peak_status="before",
+        probabilities=[{"value": 15, "probability": 0.552}],
+        shadow_probabilities=[{"value": 15, "probability": 0.324}],
+        calibration_summary={
+            "engine": "legacy",
+            "mode": "emos_shadow",
+            "calibration_version": "emos-test",
+            "calibration_source": "artifacts/probability_calibration/default.json",
+            "calibrated_mu": 15.1,
+            "calibrated_sigma": 1.25,
+        },
+        archive_path=str(archive_path),
+    )
+
+    append_probability_snapshot(**kwargs)
+    append_probability_snapshot(**kwargs)
+
+    lines = archive_path.read_text(encoding="utf-8").strip().splitlines()
+    assert len(lines) == 1
+
+
+def test_append_probability_snapshot_writes_on_bucket_change(tmp_path: Path):
+    archive_path = tmp_path / "probability_training_snapshots.jsonl"
+    base_kwargs = dict(
+        city_name="ankara",
+        local_date="2026-03-20",
+        observation_time="2026-03-20T12:00:00+03:00",
+        temp_symbol="°C",
+        raw_mu=15.2,
+        raw_sigma=1.2,
+        deb_prediction=15.4,
+        ens_data={"p10": 14.8, "median": 15.8, "p90": 17.9},
+        current_forecasts={"ECMWF": 15.8, "GFS": 14.1},
+        max_so_far=15.0,
+        peak_status="before",
+        shadow_probabilities=[{"value": 15, "probability": 0.324}],
+        calibration_summary={
+            "engine": "legacy",
+            "mode": "emos_shadow",
+            "calibration_version": "emos-test",
+            "calibration_source": "artifacts/probability_calibration/default.json",
+            "calibrated_mu": 15.1,
+            "calibrated_sigma": 1.25,
+        },
+        archive_path=str(archive_path),
+    )
+
+    append_probability_snapshot(
+        probabilities=[{"value": 15, "probability": 0.552}],
+        **base_kwargs,
+    )
+    append_probability_snapshot(
+        probabilities=[{"value": 16, "probability": 0.552}],
+        **base_kwargs,
+    )
+
+    lines = archive_path.read_text(encoding="utf-8").strip().splitlines()
+    assert len(lines) == 2
diff --git a/tests/test_probability_training_dataset.py b/tests/test_probability_training_dataset.py
new file mode 100644
index 00000000..68b4602c
--- /dev/null
+++ b/tests/test_probability_training_dataset.py
@@ -0,0 +1,45 @@
+from scripts.fit_probability_calibration import _extract_samples
+
+
+def test_extract_samples_prefers_snapshot_rows_for_same_city_day():
+    history = {
+        "ankara": {
+            "2026-03-19": {
+                "actual_high": 11.0,
+                "mu": 10.8,
+                "deb_prediction": 10.9,
+                "forecasts": {"ECMWF": 10.5, "GFS": 11.2},
+                "probability_features": {
+                    "ens_median": 10.7,
+                    "ensemble_spread": 0.8,
+                    "peak_status": "before",
+                },
+            }
+        }
+    }
+    snapshot_rows = [
+        {
+            "city": "ankara",
+            "date": "2026-03-19",
+            "timestamp": "2026-03-19T12:00:00+03:00",
+            "raw_mu": 11.2,
+            "raw_sigma": 1.1,
+            "deb_prediction": 11.0,
+            "ensemble": {"p10": 10.0, "median": 11.1, "p90": 12.2},
+            "multi_model": {"ECMWF": 10.5, "GFS": 11.2},
+            "max_so_far": 10.9,
+            "peak_status": "in_window",
+        }
+    ]
+
+    samples, filled = _extract_samples(
+        history,
+        settlement_history={},
+        snapshot_rows=snapshot_rows,
+    )
+
+    assert filled == 0
+    assert len(samples) == 1
+    assert samples[0]["sample_source"] == "snapshot"
+    assert samples[0]["raw_mu"] == 11.2
+    assert samples[0]["peak_flag"] == 0.5