feat: improve AI dialogue evaluation and capability boundaries

2026-09-19 18:58:14 -07:00
parent 335d0304b1
commit 7306dc0f85
23 changed files with 1308 additions and 452 deletions
@@ -0,0 +1,107 @@
+import 'package:flutter_test/flutter_test.dart';
+import 'package:kouyu_english/core/dialogue_decision.dart';
+import 'package:kouyu_english/core/models.dart';
+
+DialogueAiIntervention acceptedAi() => const DialogueAiIntervention(
+  accepted: true,
+  goalSatisfied: true,
+  verdict: DialogueAiVerdict.accepted,
+  explanation: '表达自然。',
+);
+
+void main() {
+  group('resolveDialogueTurnDecision', () {
+    test('requires local and AI agreement for independent scene evidence', () {
+      final decision = resolveDialogueTurnDecision(
+        local: LocalDialogueVerdict.accepted,
+        ai: acceptedAi(),
+        aiWasAttempted: true,
+        isFreeScene: true,
+        supportLevel: DialogueSupportLevel.none,
+      );
+
+      expect(decision.canAdvance, isTrue);
+      expect(decision.evidenceOutcome, EvidenceKind.independentSuccess);
+      expect(decision.validationSource, DialogueValidationSource.localAndAi);
+    });
+
+    test(
+      'AI semantic acceptance advances an unknown local expression safely',
+      () {
+        final decision = resolveDialogueTurnDecision(
+          local: LocalDialogueVerdict.uncertain,
+          ai: acceptedAi(),
+          aiWasAttempted: true,
+          isFreeScene: true,
+          supportLevel: DialogueSupportLevel.none,
+        );
+
+        expect(decision.canAdvance, isTrue);
+        expect(decision.evidenceOutcome, EvidenceKind.pending);
+        expect(decision.validationSource, DialogueValidationSource.ai);
+        expect(decision.assisted, isFalse);
+      },
+    );
+
+    test('a correction can advance but is always assisted', () {
+      final decision = resolveDialogueTurnDecision(
+        local: LocalDialogueVerdict.accepted,
+        ai: acceptedAi(),
+        aiWasAttempted: true,
+        isFreeScene: true,
+        supportLevel: DialogueSupportLevel.correction,
+      );
+
+      expect(decision.canAdvance, isTrue);
+      expect(decision.evidenceOutcome, EvidenceKind.assisted);
+      expect(decision.assisted, isTrue);
+    });
+
+    test('invalid local input never advances', () {
+      final decision = resolveDialogueTurnDecision(
+        local: LocalDialogueVerdict.rejected,
+        ai: acceptedAi(),
+        aiWasAttempted: true,
+        isFreeScene: true,
+        supportLevel: DialogueSupportLevel.none,
+      );
+
+      expect(decision.canAdvance, isFalse);
+    });
+
+    test('AI outage degrades strict local acceptance to pending evidence', () {
+      final decision = resolveDialogueTurnDecision(
+        local: LocalDialogueVerdict.accepted,
+        ai: null,
+        aiWasAttempted: true,
+        isFreeScene: true,
+        supportLevel: DialogueSupportLevel.none,
+      );
+
+      expect(decision.canAdvance, isTrue);
+      expect(decision.evidenceOutcome, EvidenceKind.pending);
+      expect(decision.validationSource, DialogueValidationSource.unavailable);
+    });
+
+    test('explicit AI correction stops a local false positive', () {
+      const correction = DialogueAiIntervention(
+        accepted: false,
+        goalSatisfied: false,
+        verdict: DialogueAiVerdict.correctable,
+        suggestion: 'My name is Mia.',
+        explanation: '这句话需要修改。',
+      );
+      final decision = resolveDialogueTurnDecision(
+        local: LocalDialogueVerdict.accepted,
+        ai: correction,
+        aiWasAttempted: true,
+        isFreeScene: true,
+        supportLevel: DialogueSupportLevel.none,
+      );
+
+      expect(decision.canAdvance, isFalse);
+      expect(decision.evidenceOutcome, EvidenceKind.pending);
+      expect(decision.validationSource, DialogueValidationSource.disagreement);
+    });
+  });
+}