diff --git a/internal/ai/patrol_ai.go b/internal/ai/patrol_ai.go
index ab89dc5f1..026e2d959 100644
--- a/internal/ai/patrol_ai.go
+++ b/internal/ai/patrol_ai.go
@@ -953,6 +953,14 @@ Every finding you report should answer "what specifically happens if the operato
 - Do NOT echo severity or category. The operator already sees "warning" or "capacity"; impact must add information.
 - Do NOT fabricate consequences when you genuinely do not know. Leave ` + "`impact`" + ` empty rather than inventing one. The frontend will render an explicit "Impact not assessed" placeholder, which is more useful than guessed copy.
 
+## Authoring Evidence (what you checked)
+
+Every finding must include the ` + "`evidence`" + ` field in patrol_report_finding. This is the trust anchor that lets the operator verify your conclusion independently.
+
+- Include the specific metric values, command outputs, log entries, or tool results that led to the finding. "ZFS pool 'data' has 7 checksum errors (checked via zpool status)" is good. "Storage issue detected" is not.
+- If you used pulse_metrics, pulse_storage, pulse_read, or any investigation tool, reference the key data points from those calls.
+- The operator should be able to look at your evidence and confirm: yes, this is real, I can see the same thing.
+
 ## Final Summary Format
 
 After completing your investigation, write a concise summary using this structure:
diff --git a/internal/ai/patrol_ai_more_test.go b/internal/ai/patrol_ai_more_test.go
index 44d6beb3a..e92f6a8b6 100644
--- a/internal/ai/patrol_ai_more_test.go
+++ b/internal/ai/patrol_ai_more_test.go
@@ -268,6 +268,23 @@ func TestGetPatrolSystemPrompt_ModeSwitch(t *testing.T) {
 	}
 }
 
+func TestGetPatrolSystemPrompt_IncludesTrustScaffoldingGuidance(t *testing.T) {
+	svc := &Service{cfg: &config.AIConfig{PatrolAutonomyLevel: config.PatrolAutonomyMonitor}}
+	ps := NewPatrolService(svc, nil)
+	prompt := ps.getPatrolSystemPrompt()
+
+	required := []string{
+		"Authoring Impact",
+		"Authoring Evidence",
+		"trust anchor",
+	}
+	for _, want := range required {
+		if !strings.Contains(prompt, want) {
+			t.Fatalf("patrol system prompt missing %q", want)
+		}
+	}
+}
+
 func TestRunAIAnalysis_EarlyErrors(t *testing.T) {
 	t.Run("nil service", func(t *testing.T) {
 		ps := NewPatrolService(nil, nil)
diff --git a/internal/ai/tools/tools_patrol.go b/internal/ai/tools/tools_patrol.go
index c2adc8fab..7bb8bb101 100644
--- a/internal/ai/tools/tools_patrol.go
+++ b/internal/ai/tools/tools_patrol.go
@@ -60,16 +60,20 @@ Returns: {"ok": true, "finding_id": "...", "is_new": true/false} on success.`,
 						Description: "Detailed description of the issue found",
 					},
 					"impact": {
-						Type:        "string",
-						Description: "Operator-facing consequence-if-ignored statement: what specifically happens, and to which workloads or jobs, if the operator does nothing. Concrete operational consequences only — do not echo severity or category. Leave empty if the consequence is genuinely unknown rather than fabricating one.",
+						Type: "string",
+						Description: "Operator-facing consequence if ignored: what happens to which workloads, jobs, or data if the operator does nothing. " +
+							"Base this on the data you gathered, not on the severity level. Most infrastructure issues have knowable consequences — " +
+							"provide this whenever the data supports it. Omit only if the consequence is truly unknowable from available information.",
 					},
 					"recommendation": {
 						Type:        "string",
-						Description: "Specific actionable recommendation for the operator",
+						Description: "Specific, actionable next step the operator should take. Include verification steps where possible.",
 					},
 					"evidence": {
-						Type:        "string",
-						Description: "Specific data/metrics/commands that support this finding",
+						Type: "string",
+						Description: "Specific data points, metric values, command outputs, or tool results that led to this finding. " +
+							"This is the trust anchor: the operator uses it to verify your conclusion independently. " +
+							"Always include the key evidence — if you gathered enough data to create this finding, you have evidence to report.",
 					},
 				},
 				Required: []string{"key", "severity", "category", "resource_id", "resource_name", "resource_type", "title", "description"},