diff --git a/docs/content/docs/api/reference/deliverability-ops.mdx b/docs/content/docs/api/reference/deliverability-ops.mdx
index 664160f2..b6599612 100644
--- a/docs/content/docs/api/reference/deliverability-ops.mdx
+++ b/docs/content/docs/api/reference/deliverability-ops.mdx
@@ -810,7 +810,7 @@ At least one of the three must carry something, and the three together are cappe
 |-------|-------------|
 | `score` | Overall deliverability score, 0 to 100, higher is safer. Grounded on `rules`, never argued against it. |
 | `verdict` | One sentence on how the email will land. |
-| `findings` | Located problems, most severe first. `severity` is `high`, `warn` or `info`; `field` is `subject` or `body`; `category` is one of `trigger_word`, `tone`, `formatting`, `links`, `structure`, `authenticity`. |
+| `findings` | Located problems, most severe first. `severity` is `high`, `warn` or `info`; `field` is `subject` or `body`, absent only when the model labelled neither and nothing in the finding could be anchored in the copy; `category` is one of `trigger_word`, `tone`, `formatting`, `links`, `structure`, `authenticity`. |
 | `findings[].text` | The fragment quoted from your copy. Absent when the finding is about the email as a whole. |
 | `suggested_subject` | A rewritten subject, or absent when the current one is fine. |
 | `improvements` | Copy-level advice with no single fragment to quote. |
diff --git a/docs/public/openapi.json b/docs/public/openapi.json
index aba8dcc4..9444378a 100644
--- a/docs/public/openapi.json
+++ b/docs/public/openapi.json
@@ -29089,7 +29089,6 @@
         "description": "One located problem the AI analysis found. `text` is verified against the submitted copy before it is returned, so it is never a fragment the caller did not send.",
         "required": [
           "severity",
-          "field",
           "issue"
         ],
         "properties": {
@@ -29103,6 +29102,7 @@
           },
           "field": {
             "type": "string",
+            "description": "Absent when the model labelled neither half and nothing in the finding could be anchored in the copy.",
             "enum": [
               "subject",
               "body"
diff --git a/internal/pkg/spamcheck/spamcheck.go b/internal/pkg/spamcheck/spamcheck.go
index ccc9cb94..2917980d 100644
--- a/internal/pkg/spamcheck/spamcheck.go
+++ b/internal/pkg/spamcheck/spamcheck.go
@@ -52,8 +52,9 @@ type Input struct {
 type Finding struct {
 	// Severity is "high", "warn" or "info".
 	Severity string `json:"severity"`
-	// Field is "subject" or "body".
-	Field string `json:"field"`
+	// Field is "subject" or "body", empty when the model labelled neither and
+	// nothing in the finding could be anchored in the copy.
+	Field string `json:"field,omitempty"`
 	// Text is the exact fragment quoted from the copy, empty when the finding
 	// is about the email as a whole.
 	Text string `json:"text,omitempty"`
@@ -407,11 +408,19 @@ func severity(s string) string {
 	}
 }
 
+// field reads the half the model named. Anything it did not clearly label comes
+// back empty rather than guessed: verify fills it in from the quote when there
+// is one, and a badge naming the wrong box on a finding nothing anchored is the
+// exact error this whole panel exists to avoid.
 func field(s string) string {
-	if strings.Contains(strings.ToLower(strings.TrimSpace(s)), "subject") {
+	switch v := strings.ToLower(strings.TrimSpace(s)); {
+	case strings.Contains(v, "subject"):
 		return warmlint.FieldSubject
+	case strings.Contains(v, "body"):
+		return warmlint.FieldBody
+	default:
+		return ""
 	}
-	return warmlint.FieldBody
 }
 
 func clamp(n int) int {
diff --git a/internal/pkg/spamcheck/spamcheck_test.go b/internal/pkg/spamcheck/spamcheck_test.go
index c47e5006..a837d860 100644
--- a/internal/pkg/spamcheck/spamcheck_test.go
+++ b/internal/pkg/spamcheck/spamcheck_test.go
@@ -317,3 +317,26 @@ func TestAnalyzeAcceptsAZeroScore(t *testing.T) {
 		t.Errorf("score = %d, want the model's own 0", res.Score)
 	}
 }
+
+// A finding the model left unlabelled must not be given a half at random. With
+// a quote, verify anchors it; without one, it carries no field rather than a
+// badge sending the writer to the wrong box.
+func TestVerifyDoesNotGuessAnUnlabelledField(t *testing.T) {
+	res := analyzed(t, `{
+	  "score": 60,
+	  "findings": [
+	    {"severity":"warn","text":"limited time offer","issue":"Anchored, so the half is known."},
+	    {"severity":"warn","issue":"Nothing to anchor and no label."}
+	  ]
+	}`, "Quick question", testBody)
+
+	if len(res.Findings) != 2 {
+		t.Fatalf("got %d findings: %+v", len(res.Findings), res.Findings)
+	}
+	if res.Findings[0].Field != "body" {
+		t.Errorf("anchored finding field = %q, want body", res.Findings[0].Field)
+	}
+	if res.Findings[1].Field != "" {
+		t.Errorf("unanchored, unlabelled finding field = %q, want empty", res.Findings[1].Field)
+	}
+}
diff --git a/web/src/lib/api/models/app/campaigns/TemplateScore.ts b/web/src/lib/api/models/app/campaigns/TemplateScore.ts
index 028e2f55..581bb5e9 100644
--- a/web/src/lib/api/models/app/campaigns/TemplateScore.ts
+++ b/web/src/lib/api/models/app/campaigns/TemplateScore.ts
@@ -42,7 +42,10 @@ export interface ScoreTemplateRequest {
 // One located problem the AI analysis found in the copy.
 export interface SpamFinding {
     severity: "high" | "warn" | "info";
-    field: TemplateField;
+    /** Absent when the model labelled neither half and nothing in the finding
+     *  could be anchored in the copy, so no badge is shown rather than one
+     *  naming the wrong box. */
+    field?: TemplateField;
     /** The exact fragment quoted from the copy, absent when the finding is
      *  about the email as a whole. Verified server-side against the template,
      *  so it is never a sentence the writer did not write. */