diff --git a/mateclaw-server/src/main/java/vip/mate/goal/model/GoalCriteriaCodec.java b/mateclaw-server/src/main/java/vip/mate/goal/model/GoalCriteriaCodec.java
index f059c6f1..470cd6f8 100644
--- a/mateclaw-server/src/main/java/vip/mate/goal/model/GoalCriteriaCodec.java
+++ b/mateclaw-server/src/main/java/vip/mate/goal/model/GoalCriteriaCodec.java
@@ -58,8 +58,8 @@ public final class GoalCriteriaCodec {
 
     /**
      * Merge a per-round verdict delta into the full checklist by id. Criteria
-     * absent from the delta are preserved unchanged; the criterion text is
-     * always kept from the existing item (the verdict never carries text).
+     * absent from the delta retain their state unless their pass lacks evidence.
+     * The criterion text is always kept from the existing item (the verdict never carries text).
      */
     public static List<GoalCriterion> merge(List<GoalCriterion> existing,
                                             List<GoalChecklistVerdict.CriterionVerdict> verdicts) {
@@ -77,18 +77,22 @@ public final class GoalCriteriaCodec {
         List<GoalCriterion> merged = new ArrayList<>(existing.size());
         for (GoalCriterion c : existing) {
             GoalChecklistVerdict.CriterionVerdict v = byId.get(c.id());
-            merged.add(v == null
-                    ? c
+            GoalCriterion candidate = v == null ? c
                     : new GoalCriterion(c.id(), c.text(), v.passed(),
-                    v.evidence() != null ? v.evidence() : ""));
+                            v.evidence() != null ? v.evidence() : "");
+            // A model boolean alone cannot satisfy even the legacy semantic
+            // checklist. This checks presence, not truth or execution provenance.
+            merged.add(candidate.passed() && !hasEvidence(candidate)
+                    ? new GoalCriterion(candidate.id(), candidate.text(), false, candidate.evidence())
+                    : candidate);
         }
         return merged;
     }
 
-    /** True only when the list is non-empty and every criterion is passed. */
+    /** True only when the list is non-empty and every criterion is passed with nonblank evidence. */
     public static boolean allPassed(List<GoalCriterion> criteria) {
         return criteria != null && !criteria.isEmpty()
-                && criteria.stream().allMatch(GoalCriterion::passed);
+                && criteria.stream().allMatch(c -> c != null && c.passed() && hasEvidence(c));
     }
 
     /** Criteria not yet passed (used for the continuation prompt + gap text). */
@@ -96,7 +100,11 @@ public final class GoalCriteriaCodec {
         if (criteria == null) {
             return List.of();
         }
-        return criteria.stream().filter(c -> !c.passed()).toList();
+        return criteria.stream().filter(c -> !c.passed() || !hasEvidence(c)).toList();
+    }
+
+    private static boolean hasEvidence(GoalCriterion criterion) {
+        return criterion.evidence() != null && !criterion.evidence().isBlank();
     }
 
     /** Reassign stable ids {@code C1..Cn} in list order. */
diff --git a/mateclaw-server/src/test/java/vip/mate/goal/model/GoalCriteriaCodecTest.java b/mateclaw-server/src/test/java/vip/mate/goal/model/GoalCriteriaCodecTest.java
index 48f82af2..44bd3ace 100644
--- a/mateclaw-server/src/test/java/vip/mate/goal/model/GoalCriteriaCodecTest.java
+++ b/mateclaw-server/src/test/java/vip/mate/goal/model/GoalCriteriaCodecTest.java
@@ -2,6 +2,9 @@ package vip.mate.goal.model;
 
 import com.fasterxml.jackson.databind.ObjectMapper;
 import org.junit.jupiter.api.Test;
+import org.junit.jupiter.params.ParameterizedTest;
+import org.junit.jupiter.params.provider.NullAndEmptySource;
+import org.junit.jupiter.params.provider.ValueSource;
 
 import java.util.List;
 
@@ -77,6 +80,23 @@ class GoalCriteriaCodecTest {
         assertFalse(merged.get(0).passed());
     }
 
+    @ParameterizedTest
+    @NullAndEmptySource
+    @ValueSource(strings = {" ", "\t\n"})
+    void blankEvidenceCannotPassNewOrPersistedCriteria(String evidence) {
+        var existing = List.of(new GoalCriterion("C1", "deliver report", false, ""));
+        var merged = GoalCriteriaCodec.merge(existing, List.of(
+                new GoalChecklistVerdict.CriterionVerdict("C1", true, evidence)));
+        assertFalse(merged.getFirst().passed());
+        assertFalse(GoalCriteriaCodec.allPassed(merged));
+        assertEquals(1, GoalCriteriaCodec.remaining(merged).size());
+
+        var persisted = List.of(new GoalCriterion("C1", "deliver report", true, evidence));
+        assertFalse(GoalCriteriaCodec.allPassed(persisted));
+        assertEquals(1, GoalCriteriaCodec.remaining(persisted).size());
+        assertFalse(GoalCriteriaCodec.merge(persisted, List.of()).getFirst().passed());
+    }
+
     // ---------- allPassed / remaining ----------
 
     @Test
diff --git a/mateclaw-server/src/test/java/vip/mate/goal/service/GoalEvaluationServiceTest.java b/mateclaw-server/src/test/java/vip/mate/goal/service/GoalEvaluationServiceTest.java
index e254b499..5b6c266b 100644
--- a/mateclaw-server/src/test/java/vip/mate/goal/service/GoalEvaluationServiceTest.java
+++ b/mateclaw-server/src/test/java/vip/mate/goal/service/GoalEvaluationServiceTest.java
@@ -94,6 +94,20 @@ class GoalEvaluationServiceTest {
         when(chatModel.call(any(Prompt.class))).thenReturn(response);
     }
 
+    @Test
+    void blankEvidenceVerdictCannotCompleteOrInflateProgress() {
+        stubChatResponse("""
+                {"criterionVerdicts":[
+                  {"id":"C1","passed":true,"evidence":""},
+                  {"id":"C2","passed":true,"evidence":null}],"summary":"done"}
+                """);
+        var result = svc.evaluate(goalWithCriteria(), List.of(), "All done");
+        assertFalse(result.completed());
+        assertEquals(0.0, result.score());
+        assertTrue(result.gap().contains("DNS configured"));
+        assertTrue(result.gap().contains("TLS enabled"));
+    }
+
     // ==================== Pre-flight guards ====================
 
     @Test