Fix interview scoring so strong answers score well

Scores were capped at ~50% even for perfect interviews because the final
score weighted skills-match (exact string overlap, almost always ~0) at
40%. Rebalance to answer-driven scoring and reduce conservative bias:

- interview_api.py / report_generator.py: weight QA answers 85% and skills
match 15% (was 40/60); bump category values (Good 0.8->0.85,
Medium 0.6->0.65). Mirrored in both so screen + PDF report agree.
- interview_engine.py: rewrite the evaluate_answer prompt with a clear
rubric, tell LUNA to use the full range and not default to "Medium",
and to judge substance not grammar (answers are speech-to-text). Output
format unchanged, so parsing is untouched.

A flawless interview now scores ~85% (more with matching skills); weak
answers still score low.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

backend/routes/interview_api.py

@@ -443,13 +443,13 @@ def interview_complete():
                         if ('excellent' in score_text) or ('5' in score_text) or ('10' in score_text):
                             qa_scores.append(1.0)
                         elif ('good' in score_text) or ('4' in score_text) or ('8' in score_text) or ('9' in score_text):
-                            qa_scores.append(0.8)
+                            qa_scores.append(0.85)
                         elif ('satisfactory' in score_text) or ('medium' in score_text) or ('3' in score_text) or ('6' in score_text) or ('7' in score_text):
-                            qa_scores.append(0.6)
+                            qa_scores.append(0.65)
                         elif ('needs improvement' in score_text) or ('poor' in score_text) or ('2' in score_text):
                             qa_scores.append(0.4)
                         else:
-                            qa_scores.append(0.2)
+                            qa_scores.append(0.25)
             except Exception:
                 qa_scores = []
 
@@ -459,10 +459,11 @@ def interview_complete():
             # data.
             qa_average = (sum(qa_scores) / len(qa_scores)) if qa_scores else 0.5
 
-            # Weight skills match (40%) and QA average (60%) to derive
-            # the final overall score.  Convert to a percentage for
-            # display.
-            overall = (ratio * 0.4) + (qa_average * 0.6)
+            # Weight the interview answers (85%) as the primary signal and the
+            # skills match (15%) as a minor bonus.  The skills ratio relies on
+            # exact skill-string overlap, which is often near zero even for
+            # strong candidates, so it must not dominate the final score.
+            overall = (qa_average * 0.85) + (ratio * 0.15)
             percentage = overall * 100.0
 
             # Assign a descriptive label based on the overall score.

backend/services/interview_engine.py

@@ -475,17 +475,24 @@ def evaluate_answer(question, answer, job_role="Software Developer", seniority="
             }
             
         prompt = f"""
-        You are LUNA, an experienced recruiter evaluating a candidate's spoken answer for a
-        {seniority} {job_role} position. Judge it at the level expected for that seniority.
+        You are LUNA, an experienced technical recruiter evaluating a candidate's spoken answer
+        for a {seniority} {job_role} position.
 
         Question: {question}
         Candidate Answer: {answer}
 
-        Evaluate on: technical correctness, relevance to the question, depth/specificity, and clarity.
-        Calibrate to seniority — be more demanding for senior roles, more forgiving for junior ones.
-        Reward concrete, specific answers; penalise vague, off-topic, or empty ones.
-
-        Rate the answer as exactly one of: Poor, Medium, Good, Excellent.
+        Rate the answer using this rubric:
+        - Excellent: correct, relevant and specific; shows clear understanding appropriate for a
+          {seniority} {job_role}. Small imperfections are fine.
+        - Good: mostly correct and relevant with reasonable detail, but missing some depth or specifics.
+        - Medium: partially correct or relevant, but vague, generic or incomplete.
+        - Poor: incorrect, off-topic, or no real answer.
+
+        Important:
+        - Judge the SUBSTANCE of the answer, not its grammar or wording — it was transcribed from
+          speech, so ignore filler words, punctuation and minor phrasing issues.
+        - A clear, correct, on-topic answer should be rated "Good" or "Excellent". Do NOT default to
+          "Medium" — use the full range and reward strong answers.
 
         Respond in this exact format and nothing else:
         Score: [Poor/Medium/Good/Excellent]

backend/services/report_generator.py

@@ -476,18 +476,19 @@ def _calculate_overall_score(report_data: Dict[str, Any]) -> Dict[str, Any]:
         if 'excellent' in score_text or '5' in score_text or '10' in score_text:
             qa_scores.append(1.0)
         elif 'good' in score_text or '4' in score_text or '8' in score_text or '9' in score_text:
-            qa_scores.append(0.8)
+            qa_scores.append(0.85)
         elif 'satisfactory' in score_text or 'medium' in score_text or '3' in score_text or '6' in score_text or '7' in score_text:
-            qa_scores.append(0.6)
-        elif 'needs improvement' in score_text or 'poor' in score_text or '2' in score_text or '4' in score_text or '5' in score_text:
+            qa_scores.append(0.65)
+        elif 'needs improvement' in score_text or 'poor' in score_text or '2' in score_text:
             qa_scores.append(0.4)
         else:
-            qa_scores.append(0.2)
-    
+            qa_scores.append(0.25)
+
     qa_average = sum(qa_scores) / len(qa_scores) if qa_scores else 0.5
-    
-    # Calculate weighted average
-    overall = (skills_ratio * 0.4) + (qa_average * 0.6)
+
+    # Weight the interview answers (85%) as the primary signal and the skills
+    # match (15%) as a minor bonus, mirroring backend/routes/interview_api.py.
+    overall = (qa_average * 0.85) + (skills_ratio * 0.15)
     percentage = overall * 100
     
     if overall >= 0.8: