Update README.md

README.md

@@ -62,12 +62,22 @@ Evaluated on 70 held-out Taiwan-style test cases:
 | JSON format accuracy | **100%** |
 | PII judgment accuracy | **98.6%** |
 | Entity recall | **100%** |
-| Mask accuracy (with rule-based post-processing) | **~100%** |
+| Mask accuracy (with included safeguards) | **~100%** |
 
 The model produces correct PII values in 100% of cases. The included
 post-processing (`validate_and_fix_mask`) recomputes masks deterministically
 from the value, ensuring output masks are always character-accurate.
 
+## Inference safeguards
+
+The included `inference.py` applies two filters and one mask corrector:
+
+- **Structural validation** — drops entities whose value shape doesn't match the claimed type (e.g., `BANK_ACCOUNT` value containing non-digits, or `ADDRESS` value lacking "號" or "樓")
+- **Provenance check** — drops entities whose value doesn't verbatim appear in the input (rejects hallucinations and mistranslations)
+- **Mask correction** — recomputes the `masked` field deterministically from `value`, guaranteeing correct star count
+
+These guarantee output integrity even when the model occasionally misclassifies or fabricates.
+
 ## Quick start
 
 ```bash
@@ -174,12 +184,22 @@ correspond to any real individuals.
 | JSON 格式正確率 | **100%** |
 | PII 判斷正確率 | **98.6%** |
 | 實體召回率 | **100%** |
-| 遮罩正確率（含規則式後處理） | **~100%** |
+| 遮罩正確率（含內建安全機制） | **~100%** |
 
 模型在所有測試案例中皆能正確抽出 PII 原始值。內附的後處理函式
 （`validate_and_fix_mask`）會根據抽出的值重新計算遮罩，確保星號數量
 永遠正確。
 
+## 推論安全機制
+
+內附的 `inference.py` 套用兩道過濾與一道遮罩修正：
+
+- **結構驗證** — 丟棄 value 結構不符其宣稱類型的 entity（例如 `BANK_ACCOUNT` 含非數字字元、或 `ADDRESS` 缺「號」或「樓」）
+- **來源驗證** — 丟棄 value 未逐字出現在原文中的 entity（擋幻覺與翻譯錯誤）
+- **遮罩修正** — 從 value 確定性地重算 masked 欄位，保證星號數量精準
+
+這幾層確保即使模型偶有誤判或捏造，輸出仍然可信。
+
 ## 快速使用
 
 ```bash