diff --git a/go/internal/api/routes/core.go b/go/internal/api/routes/core.go
index 880eaa5..d34ba91 100644
--- a/go/internal/api/routes/core.go
+++ b/go/internal/api/routes/core.go
@@ -48,8 +48,29 @@ func respond(w http.ResponseWriter, code int, data interface{}) {
 	json.NewEncoder(w).Encode(data)
 }
 
+// Unified error body schema (t_2d4e3b92, 2026-09-21):
+//
+//	HTTP non-2xx  →  {"error":{"code":<string>,"message":<string>,"retry_after_s":<int>}}
+//	HTTP 2xx      →  success schema only (recall: {"results":[...],"count":N,...})
+//
+// Contract for callers: 200 + "results" key present = success (results MAY be empty —
+// a legitimate "no match"); non-2xx = failure (never parse as success). Do NOT
+// distinguish by len(results)==0.
 func respondError(w http.ResponseWriter, code int, msg string) {
-	respond(w, code, map[string]string{"error": msg})
+	respond(w, code, map[string]interface{}{"error": map[string]interface{}{
+		"code":          "failure",
+		"message":       msg,
+		"retry_after_s": 0,
+	}})
+}
+
+// respondErrorSchema returns a non-2xx with the unified error schema.
+func respondErrorSchema(w http.ResponseWriter, code int, codeStr, msg string, retryAfterS int) {
+	respond(w, code, map[string]interface{}{"error": map[string]interface{}{
+		"code":          codeStr,
+		"message":       msg,
+		"retry_after_s": retryAfterS,
+	}})
 }
 
 // POST /api/v1/commit
@@ -273,11 +294,27 @@ func (a *API) Recall(w http.ResponseWriter, r *http.Request) {
 	switch req.Mode {
 	case "keyword":
 		// Pure keyword search: LanceDB + BM25 re-rank, or graph.db fallback
+		// t_2d4e3b92: keyword 主路径依赖向量编码；编码失败（依赖不可用）⇒ 非 2xx，
+		// 不再把错误伪装成 200 + 错误体。
 		kwResults, kwErr := a.Pipeline.Recall(req.Query, req.Namespace, req.Limit, req.Diversity)
-		if kwErr != nil || len(kwResults) == 0 {
+		if kwErr != nil {
+			// 主路径硬错误：兜底命中也按"主路径故障"报告（502），调用方据此走失败路径
+			if fb := a.GraphStore.FallbackTextSearch(req.Query, req.Namespace, req.Limit); len(fb) > 0 {
+				_ = fb
+			}
+			respondErrorSchema(w, 502, "recall_backend_error",
+				"keyword recall failed: "+kwErr.Error(), 5)
+			log.Printf("[recall:keyword] failed: %v", kwErr)
+			return
+		}
+		if len(kwResults) == 0 {
+			// 真无匹配（正常业务）：成功 schema，results=[] + candidates_scanned 供区分
 			fallbackResults := a.GraphStore.FallbackTextSearch(req.Query, req.Namespace, req.Limit)
 			results = convertFallbackResults(fallbackResults)
-			respond(w, 200, map[string]interface{}{"results": results, "count": len(results), "mode": "keyword"})
+			respond(w, 200, map[string]interface{}{
+				"results": results, "count": len(results), "mode": "keyword",
+				"candidates_scanned": 0, // keyword 路径未做 ANN 扫描
+			})
 			return
 		}
 		// Re-rank by BM25 only
@@ -298,13 +335,16 @@ func (a *API) Recall(w http.ResponseWriter, r *http.Request) {
 		results, err = a.Pipeline.Recall(req.Query, req.Namespace, req.Limit, req.Diversity)
 	}
 	if err != nil {
-		// P0: 降级到 graph.db 关键词搜索
+		// P0: 降级到 graph.db 关键词搜索（仅 hybrid/semantic 主路径依赖向量/rerank）
 		fallbackResults := a.GraphStore.FallbackTextSearch(req.Query, req.Namespace, req.Limit)
 		if len(fallbackResults) > 0 {
 			results = convertFallbackResults(fallbackResults)
 			w.Header().Set("X-Fallback", "graph")
 		} else {
-			respondError(w, 500, "recall failed: "+err.Error())
+			// t_2d4e3b92: 主路径与兜底都失败 ⇒ 非 2xx + 统一错误体（绝不在 200 里塞错误）
+			respondErrorSchema(w, 502, "recall_backend_error",
+				"recall failed: "+err.Error()+"; graph fallback also empty", 5)
+			log.Printf("[recall] failed: %v; fallback empty", err)
 			return
 		}
 	}
diff --git a/go/internal/api/routes/eval.go b/go/internal/api/routes/eval.go
index 6041cfd..78be494 100644
--- a/go/internal/api/routes/eval.go
+++ b/go/internal/api/routes/eval.go
@@ -4,11 +4,130 @@ package routes
 import (
 	"encoding/json"
 	"net/http"
+	"os"
+	"strings"
 	"sync"
 	"time"
 
 	"github.com/xiaoxue/memoryweave/internal/storage"
 )
+
+// ─── 金标集定义 ──────────────────────────────────────────────
+
+// GoldenSetItem 单条金标：query + 具体 expected_ids + 关键词相关性判据
+type GoldenSetItem struct {
+	Query              string   `json:"query"`
+	ExpectedIDs        []string `json:"expected_ids"`
+	MustIncludeKeywords []string `json:"must_include_keywords"`
+	Category           string   `json:"category,omitempty"`
+	Difficulty         string   `json:"difficulty,omitempty"`
+	IDBaseline         string   `json:"id_baseline,omitempty"`
+}
+
+// GoldenSet 金标集
+type GoldenSet struct {
+	Description string          `json:"description"`
+	Items       []GoldenSetItem `json:"items"`
+}
+
+// LoadGoldenSet 从 testdata/ 加载金标集，支持相对路径（生产用）和绝对路径（测试用）
+func LoadGoldenSet(path string) (*GoldenSet, error) {
+	if path == "" {
+		// 编译后无法直接拿到源码路径，回退到 CWD 的 testdata/
+		path = "testdata/golden-set.json"
+	}
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return nil, err
+	}
+	var gs GoldenSet
+	if err := json.Unmarshal(data, &gs); err != nil {
+		return nil, err
+	}
+	// 空断言自守：禁止裸前缀/空串混入（避免恒真）
+	for i, item := range gs.Items {
+		if err := validateExpectedIDs(item.Query, item.ExpectedIDs); err != nil {
+			return nil, err
+		}
+		_ = i
+	}
+	return &gs, nil
+}
+
+// validateExpectedIDs 空断言自守：裸前缀 "mem_"、空串、None → 直接报错，不许静默 PASS
+func validateExpectedIDs(query string, expectedIDs []string) error {
+	if len(expectedIDs) == 0 {
+		return nil // 允许空（bootstrap 时还没基线）
+	}
+	for _, id := range expectedIDs {
+		trimmed := strings.TrimSpace(id)
+		// 裸前缀 "mem_"（不含具体 ID 数字）
+		if trimmed == "mem_" || trimmed == "mem" {
+			return &VacuousAssertionError{
+				Query:  query,
+				BadID:  id,
+				Reason: "裸前缀 mem_ 是恒真断言——所有记忆 ID 都以 mem_ 开头",
+			}
+		}
+		// 空串
+		if trimmed == "" {
+			return &VacuousAssertionError{
+				Query:  query,
+				BadID:  id,
+				Reason: "空串期望值（无意义断言）",
+			}
+		}
+		// None（JSON null 解码成空字符串已处理，但检查 "None" 字符串）
+		if strings.EqualFold(trimmed, "none") {
+			return &VacuousAssertionError{
+				Query:  query,
+				BadID:  id,
+				Reason: "None 期望值（无意义断言）",
+			}
+		}
+	}
+	return nil
+}
+
+// VacuousAssertionError 空断言错误
+type VacuousAssertionError struct {
+	Query  string
+	BadID  string
+	Reason string
+}
+
+func (e *VacuousAssertionError) Error() string {
+	return "FATAL 空断言被检出（恒真）: query=" + e.Query + " 期望=" + e.BadID + " — " + e.Reason
+}
+
+// checkKeywordRelevance 关键词相关性判据：返回内容必须包含至少一个关键词
+func checkKeywordRelevance(content string, keywords []string) bool {
+	if len(keywords) == 0 {
+		// 无关键词判据 → 视为不通过（禁止退化为恒真）
+		return false
+	}
+	lower := strings.ToLower(content)
+	for _, kw := range keywords {
+		if strings.Contains(lower, strings.ToLower(kw)) {
+			return true
+		}
+	}
+	return false
+}
+
+// validateGoldenSetForRun 对一次 eval run 的所有 expected_ids 做空断言自守
+func validateGoldenSetForRun(queries []struct {
+	Query       string   `json:"query"`
+	ExpectedIDs []string `json:"expected_ids"`
+}) error {
+	for _, q := range queries {
+		if err := validateExpectedIDs(q.Query, q.ExpectedIDs); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
 type EvalAPI struct {
 	Pipeline *storage.RecallPipeline
 	LanceDB  storage.LanceDB
@@ -39,13 +158,24 @@ type EvalRun struct {
 	RanAt   time.Time `json:"ran_at"`
 }
 
+// EvalQueryDetail
 type EvalQueryDetail struct {
-	Query       string   `json:"query"`
-	ExpectedIDs []string `json:"expected_ids"`
-	Hits        []string `json:"hits"`
-	Misses      []string `json:"misses"`
-	RecallAt5   float64  `json:"recall_at_5"`
-	PrecisionAt5 float64 `json:"precision_at_5"`
+	Query             string   `json:"query"`
+	ExpectedIDs       []string `json:"expected_ids"`
+	MustIncludeKeywords []string `json:"must_include_keywords,omitempty"`
+	Hits              []string `json:"hits"`
+	Misses            []string `json:"misses"`
+	RecallAt5         float64  `json:"recall_at_5"`
+	PrecisionAt5      float64  `json:"precision_at_5"`
+	KeywordOK         bool     `json:"keyword_ok,omitempty"`
+}
+
+// GoldenRunResult 金标 run 的扩展结果（含关键词判据）
+type GoldenRunResult struct {
+	EvalRun
+	KeywordPass int `json:"keyword_pass"`
+	KeywordFail int `json:"keyword_fail"`
+	KeywordTotal int `json:"keyword_total"`
 }
 
 type ConsolidationHits struct {
diff --git a/go/internal/api/routes/testdata/golden-set.json b/go/internal/api/routes/testdata/golden-set.json
new file mode 100644
index 0000000..e68124a
--- /dev/null
+++ b/go/internal/api/routes/testdata/golden-set.json
@@ -0,0 +1,117 @@
+{
+  "description": "织忆召回金标检查集（12条基线 + 2条目标知识）",
+  "items": [
+    {
+      "query": "织忆是什么",
+      "expected_ids": ["mem_1789153726378940805"],
+      "must_include_keywords": ["织忆"],
+      "category": "system_fact",
+      "difficulty": "easy",
+      "id_baseline": "mem_1789153726378940805"
+    },
+    {
+      "query": "织忆的图谱扩展机制是什么",
+      "expected_ids": ["mem_1789823837902661249"],
+      "must_include_keywords": ["图谱"],
+      "category": "system_fact",
+      "difficulty": "medium",
+      "id_baseline": "mem_1789823837902661249"
+    },
+    {
+      "query": "织忆如何通过 E1 提取和 MMR 实现多样去重",
+      "expected_ids": ["mem_1789915227303612056"],
+      "must_include_keywords": ["MMR"],
+      "category": "system_fact",
+      "difficulty": "hard",
+      "id_baseline": "mem_1789915227303612056"
+    },
+    {
+      "query": "牧尘的偏好是什么",
+      "expected_ids": ["mem_1789667544872361813"],
+      "must_include_keywords": ["牧尘"],
+      "category": "user_pref",
+      "difficulty": "easy",
+      "id_baseline": "mem_1789667544872361813"
+    },
+    {
+      "query": "牧尘喜欢什么样的工作方式",
+      "expected_ids": ["mem_1789837840173161705"],
+      "must_include_keywords": ["牧尘"],
+      "category": "user_pref",
+      "difficulty": "medium",
+      "id_baseline": "mem_1789837840173161705"
+    },
+    {
+      "query": "如何根据牧尘的偏好调整记忆检索策略",
+      "expected_ids": ["mem_1789190217889697697"],
+      "must_include_keywords": ["记忆"],
+      "category": "user_pref",
+      "difficulty": "hard",
+      "id_baseline": "mem_1789190217889697697"
+    },
+    {
+      "query": "当前项目的技术栈是什么",
+      "expected_ids": ["mem_1789680870868739202"],
+      "must_include_keywords": ["技术栈"],
+      "category": "proj_context",
+      "difficulty": "easy",
+      "id_baseline": "mem_1789680870868739202"
+    },
+    {
+      "query": "织忆项目有哪些核心组件",
+      "expected_ids": ["mem_1789154058288011112"],
+      "must_include_keywords": ["织忆"],
+      "category": "proj_context",
+      "difficulty": "medium",
+      "id_baseline": "mem_1789154058288011112"
+    },
+    {
+      "query": "织忆和其他记忆系统相比有什么架构优势",
+      "expected_ids": ["mem_1789153726378940805"],
+      "must_include_keywords": ["织忆"],
+      "category": "proj_context",
+      "difficulty": "hard",
+      "id_baseline": "mem_1789153726378940805"
+    },
+    {
+      "query": "如何使用 recall 接口",
+      "expected_ids": ["mem_1787272787836687543"],
+      "must_include_keywords": ["recall"],
+      "category": "tool_usage",
+      "difficulty": "easy",
+      "id_baseline": "mem_1787272787836687543"
+    },
+    {
+      "query": "织忆有哪些管理工具",
+      "expected_ids": ["mem_1789153726378940805"],
+      "must_include_keywords": ["织忆"],
+      "category": "tool_usage",
+      "difficulty": "medium",
+      "id_baseline": "mem_1789153726378940805"
+    },
+    {
+      "query": "如何通过 API 扩展织忆功能",
+      "expected_ids": ["mem_1789153726378940805"],
+      "must_include_keywords": ["织忆"],
+      "category": "tool_usage",
+      "difficulty": "hard",
+      "id_baseline": "mem_1789153726378940805"
+    },
+    {
+      "query": "本地 llama 服务 8080 端口",
+      "expected_ids": ["mem_1789229548476546193"],
+      "must_include_keywords": ["8080"],
+      "category": "target_knowledge",
+      "difficulty": "easy",
+      "id_baseline": "mem_1789229548476546193"
+    },
+    {
+      "query": "session-to-zhiyi.py 脚本名",
+      "expected_ids": ["mem_1788977361686058470"],
+      "must_include_keywords": ["session-to-zhiyi"],
+      "category": "target_knowledge",
+      "difficulty": "easy",
+      "id_baseline": "mem_1788977361686058470"
+    }
+  ]
+}