Kalibrierte Scan-Confidence + Cross-Model-Bonus

__tests__/server/scanReview.test.js

@@ -82,6 +82,35 @@ describe('decideReviewOutcome', () => {
     expect(decision.reason).toBe('review-confirmed-primary');
   });
 
+  it('boosts confidence when two different models agree on the species', () => {
+    // The 55%-forever bug: primary 0.55 + agreeing review 0.54 must NOT stay
+    // at 0.55 — independent cross-model agreement is real evidence.
+    const decision = decideReviewOutcome({
+      primaryResult: ai('Sacred lotus', 'Nelumbo nucifera', 0.55),
+      reviewResult: ai('Sacred lotus', 'Nelumbo nucifera', 0.54),
+      agrees: true,
+    });
+    expect(decision.confidence).toBeCloseTo(0.75, 5);
+  });
+
+  it('caps the agreement-boosted confidence at 0.97', () => {
+    const decision = decideReviewOutcome({
+      primaryResult: ai('Sacred lotus', 'Nelumbo nucifera', 0.95),
+      reviewResult: ai('Sacred lotus', 'Nelumbo nucifera', 0.9),
+      agrees: true,
+    });
+    expect(decision.confidence).toBe(0.97);
+  });
+
+  it('gives no confidence bonus on disagreement', () => {
+    const decision = decideReviewOutcome({
+      primaryResult: ai('Rose', 'Rosa chinensis', 0.5),
+      reviewResult: ai('Sacred lotus', 'Nelumbo nucifera', 0.6),
+      agrees: false,
+    });
+    expect(decision.confidence).toBeUndefined();
+  });
+
   it('replaces with the agreeing review on a confidence tie (stronger model wins ties)', () => {
     const decision = decideReviewOutcome({
       primaryResult: ai('Swiss Cheese Plant', 'Monstera deliciosa', 0.7),

app.json

@@ -2,7 +2,7 @@
   "expo": {
     "name": "GreenLens",
     "slug": "greenlens",
-    "version": "2.2.8",
+    "version": "2.2.9",
     "orientation": "portrait",
     "icon": "./assets/icon.png",
     "userInterfaceStyle": "automatic",

package.json

@@ -1,6 +1,6 @@
 {
   "name": "greenlens",
-  "version": "2.2.7",
+  "version": "2.2.9",
   "main": "expo-router/entry",
   "private": true,
   "scripts": {

server/index.js

@@ -822,6 +822,10 @@ app.post('/v1/scan', async (request, response) => {
                 modelUsed = openAiReview.modelUsed || modelUsed;
                 if (grounded.grounded) modelPath.push('catalog-grounded-review');
               }
+              if (decision.confidence != null) {
+                // Cross-model agreement bonus: both models named the same species.
+                result = { ...result, confidence: decision.confidence };
+              }
               modelPath.push('openai-review');
               modelPath.push(decision.reason);
             } else {

server/lib/openai.js

@@ -243,8 +243,11 @@ const buildIdentifyPrompt = (language, mode) => {
       : '- "careInfo.temp": temperature range in Celsius (e.g. "18–24 °C"). Must always be a real plant-specific value, never "Unknown".',
     '- "botanicalName" must use accepted Latin scientific naming and must not be invented or misspelled.',
     '- If species is uncertain, prefer genus-level naming (for example: "Calathea sp.").',
-    '- "confidence" must be between 0 and 1.',
-    '- Keep confidence <= 0.55 when the image is ambiguous, blurred, or partially visible.',
+    '- "confidence" must be between 0 and 1 and reflect how certain the species identification is. Calibrate it:',
+    '  - 0.85-0.95: species clearly recognizable, distinctive features (leaf shape, flower, pattern) plainly visible.',
+    '  - 0.65-0.84: species very likely, but some distinguishing features are hidden or similar species exist.',
+    '  - 0.40-0.64: image is ambiguous, blurred, partially visible, or several species fit equally well.',
+    '  - Below 0.40: mostly guessing; prefer genus-level naming instead.',
     '- "waterIntervalDays" must be an integer between 1 and 45.',
     '- Do not include markdown, explanations, or extra keys.',
   ].join('\n');

server/lib/scanReview.js

@@ -1,5 +1,13 @@
 const { normalizeText } = require('./scanGrounding');
 
+const clamp = (value, min, max) => Math.min(max, Math.max(min, value));
+
+// Two DIFFERENT models independently naming the same species is genuine
+// evidence beyond either model's self-reported confidence, so agreement
+// earns a bonus on top of the better single estimate.
+const REVIEW_AGREEMENT_BONUS = 0.2;
+const REVIEW_AGREEMENT_CONFIDENCE_CAP = 0.97;
+
 // Agreement must be judged on the RAW model answers, not the grounded ones:
 // catalog grounding has a genus-level fallback that can collapse two different
 // species onto the same catalog entry and fake an agreement.
@@ -30,6 +38,11 @@ const decideReviewOutcome = ({ primaryResult, reviewResult, agrees }) => {
       accept: true,
       replace: reviewConfidence >= primaryConfidence,
       reason: 'review-confirmed-primary',
+      confidence: clamp(
+        Math.max(primaryConfidence, reviewConfidence) + REVIEW_AGREEMENT_BONUS,
+        0.05,
+        REVIEW_AGREEMENT_CONFIDENCE_CAP,
+      ),
     };
   }