Update KAGGLE_SUBMIT.md with current status and v4 instructions

medal-solvers/KAGGLE_SUBMIT.md

@@ -1,6 +1,6 @@
 # Kaggle Submission Instructions
 
-## Current Status: LB 6053.17 (V17), Bronze threshold: 6055.14
+## Current Status: LB 6059.08 (V25), Target: 6100
 
 ## How to Submit (Kaggle Notebook)
 
@@ -25,26 +25,40 @@ os.chdir('/kaggle/working/neurogolf-solver/medal-solvers')
 ## What swap_and_submit.py does:
 1. Opens submission-6043.zip (base submission)
 2. Auto-discovers ALL .onnx files in `optimized/` folder:
-   - `optimized/task255.onnx` (11.570 pts, +4.93 from original)
-   - `optimized/task285.onnx` (8.818 pts, +2.83 from original)  
-   - `optimized/task319.onnx` (11.244 pts, +3.32 from original)
+   - `optimized/task084.onnx` (~12 pts, +2.56 from base)
+   - `optimized/task153.onnx` (**V4: 14.0 pts local**, fixes Kaggle discrepancy)
+   - `optimized/task255.onnx` (11.570 pts, +4.93 from base)
+   - `optimized/task319.onnx` (11.580 pts, +3.66 from base)
 3. Validates each model against task data (0 failures required)
 4. Creates new submission.zip with replaced models
 
-## Score Budget
+## Task 153 V4 — Key Change
 
-| Task | Original | Optimized | Gain | Status |
-|------|----------|-----------|------|--------|
-| 285 | 5.98 | 8.82 | +2.83 | ✅ Confirmed on Kaggle |
-| 255 | 6.64 | 11.57 | +4.93 | ✅ In submission |
-| 319 | 7.92 | 11.24 | +3.32 | ✅ In submission |
-| **Total** | | | **+11.08** | **LB: 6053.17** |
+V4 eliminates the multi-dimensional Gather `[9,3,3]` pattern that caused 
+the Kaggle scoring discrepancy (13.83 local vs 7.92 Kaggle in V3).
 
-## Gap to Bronze: 1.97 pts
+**Expected impact**: If Kaggle matches local, task153 goes from ~7.92 → ~14 = **+6 pts**
 
-To close the gap, would need to optimize another task model. 
-Largest models with most optimization potential:
-- task209.onnx (1.29MB), task366.onnx (1.27MB), task084.onnx (1.13MB)
-- task076.onnx (949KB), task233.onnx (939KB), task153.onnx (862KB)
+## Score Budget (with V4)
 
-Each would require reverse-engineering the task rule (like we did for 255/285/319).
+| Task | Base | V25 LB | V4 Expected | Expected Gain vs V25 |
+|------|------|--------|-------------|---------------------|
+| 084 | ~9.4 | ~12 | ~12 | — |
+| 153 | 4.91 | ~7.92 | ~14.0 | **+6.0** |
+| 255 | 6.64 | 11.57 | 11.57 | — |
+| 285 | 5.98 | 8.82 | 8.82 | — |
+| 319 | 7.92 | 11.58 | 11.58 | — |
+| **Total Expected** | | **6059.08** | **~6065** | **+6** |
+
+## Remaining Gap After V4: ~35 pts to 6100
+
+Need to optimize 5-7 more tasks at +5 pts each to reach 6100.
+Priority targets: task366, task025, task158, task118, task209.
+
+## Important Notes
+
+- **task285.onnx** is NOT in the optimized/ folder — it was already included 
+  in the original V13 submission. The base zip (submission-6043.zip) already 
+  uses the optimized task285.
+- Always validate before submitting: a model with ANY failures scores only 1.0!
+- File size limit per model: 1,509,949 bytes