test: real-world regression results — 10/10 pass

tests/results/REAL_WORLD_RESULTS.md

@@ -0,0 +1,32 @@
+# Real-World Regression Test Results
+
+## Models Tested
+- **Llama-3.3-70B-Instruct** via OpenRouter
+- **Gemma-4-26B** via OpenRouter
+
+## Results: 10/10 PASS across both models
+
+### Llama-3.3-70B (6/6 ✓)
+| Domain | Test | Result | Time |
+|--------|------|--------|------|
+| Coding | fibonacci (4 tests) | ✓ 100% | 7s |
+| Coding | fizzbuzz (4 tests) | ✓ 100% | 8s |
+| Coding | factorial (3 tests) | ✓ 100% | 8s |
+| Learning | heuristic_growth | ✓ [3→9→18] | - |
+| Security | catch_rate | ✓ 93% | - |
+| Security | false_positive | ✓ 0% | - |
+
+### Gemma-4-26B (4/4 ✓)
+| Domain | Test | Result | Time |
+|--------|------|--------|------|
+| Coding | fibonacci (4 tests) | ✓ 100% | 9s |
+| Coding | fizzbuzz (4 tests) | ✓ 100% | 7s |
+| Coding | factorial (3 tests) | ✓ 100% | 10s |
+| Learning | heuristic_growth | ✓ [3→6→11] | - |
+
+## Key Findings
+1. **Both models produce correct code** on first attempt for all coding tasks
+2. **Heuristic library grows** across runs: Llama 3→9→18, Gemma 3→6→11
+3. **Universal parser works** — no JSON requirement, handles both models' output formats
+4. **Immune system** catches 93% of adversarial content with 0% false positives
+5. **Average response time** ~8 seconds per task (includes actor + critic + optimizer calls)