Update TODO.md: deprioritize 023, focus on 133/158/018 for 14pts bronze gap

medal-solvers/TODO.md

@@ -1,98 +1,63 @@
 # medal-solvers — TODO (ORDERED)
 
-## Current: 6139.43 (V119) | Target: 6500 (silver)
+## Current: ~6139 (V119) | Target: 6153+ (bronze) | Gap: 14 pts
 
 ---
 
-## ⭐⭐⭐ #1 PRIORITY: TASK 023 — REDUCED ONNX IMPLEMENTATION ⭐⭐⭐
-
-### Current: score=11.06 (370 nodes, 1.33M memory) | Second dataset: 11.62
-### Target: 13-15+ if reduced algorithm works
-
-### ALGORITHM IS KNOWN (reverse-engineered from 370-node model):
-The algorithm uses **iterative endpoint propagation** — 12 iterations of:
-1. Compute MinPool(2×2), H-triplet(1×3), V-triplet(3×1)
-2. Convolve with shifted kernels, find cells with exactly 1 evidence
-3. Intersect with previous features → refined markers
-4. After all iterations: body = remaining MinPool, edge = remaining triplets
-
-### Current bottleneck: 370 nodes create ~369 intermediates of [1,1,30,30] = 1.33M memory.
-
-### Path to higher score:
-1. **FIX Python reimplementation** (currently 76.74% — threshold bugs at iteration step)
-   - The model uses threshold 3.5/2.5 for initial extraction, 0.5 for subsequent intersections
-   - Need to trace exact values at each step to match
-2. **Reduce iterations** using LARGER kernels:
-   - 5×5 kernel reaches 2 cells per iteration → 6 iterations instead of 12
-   - 7×7 kernel reaches 3 cells → 4 iterations
-   - Each iteration has ~30 nodes. Reducing from 12→6 saves ~180 nodes = ~180 intermediates
-   - Memory drops from 1.33M to ~650K → score ~11.6 → to ~320K → score ~12.3
-3. **Build reduced ONNX** and validate on all 266 + random generated examples
-4. **Profile on Kaggle** (static scoring is UNRELIABLE)
-
-### DO NOT:
-- ❌ Don't try gradient-trained approach (Kaggle rejected it — doesn't generalize)
-- ❌ Don't use static scoring to estimate gains
-- ❌ Don't declare the problem impossible — the algorithm IS known and works
-
-### Training script (backup plan):
-`medal-solvers/build_tasks/train_task023.py` — run on Kaggle GPU with N_GEN=5000+.
-Uses oracle (370-node model) to label random shapes. If hand-crafted approach stalls, 
-this is the fallback but MUST verify on 5000+ random shapes before submitting.
-
----
-
-## ✅ DONE: TASK 258 — Score 18.187 → 19.925 (+1.738, V118)
-Single Conv[10,10,1,3] with bias threshold AND trick. See `convolution_series_part_2.py`.
-
-## ✅ DONE: TASK 285 — Score 8.836 → 10.696 (+1.86 pts, V114)
-
----
-
-## ⭐⭐ #2: TASK 133 — PER-CHANNEL CONVERSION ⭐⭐
+## ⭐⭐⭐ #1 PRIORITY: TASK 133 — PER-CHANNEL CONVERSION ⭐⭐⭐
 
 ### Current: score=8.97, memory=9,115,804
 ### Target: score=11-12 (gain +2-3 pts)
 
 Unroll per-color: process at [1,1,30,30] instead of [1,9,30,30].
+Same technique as task 025. See `build_task025_onnx.py`.
 
 ---
 
-## ⭐⭐ #3: TASK 158 — PROPAGATION REBUILD ⭐⭐
+## ⭐⭐⭐ #2: TASK 158 — PROPAGATION REBUILD ⭐⭐⭐
 
 ### Current: score=9.30, memory=6,563,608
 ### Target: score=11-12 (gain +1.5-2.5 pts)
 
-MaxPool propagation framework (same as task 285 V4).
+MaxPool propagation framework (same as task 285 V4 which gave +1.86).
+See `build_task285_v4_propagation.py`.
 
 ---
 
-## ⭐ #4: TASK 018 — NODE REDUCTION ⭐
+## ⭐⭐ #3: TASK 018 — NODE REDUCTION ⭐⭐
 
 ### Current: score=11.32, memory=868,894
 ### Target: score=12.5-13.5 (gain +1.2-2.2 pts)
 
 ---
 
-## ⭐ #5: TASK 255 — ELIMINATE [30,30] MATRICES ⭐
+## ⭐⭐ #4: TASK 255 — ELIMINATE [30,30] MATRICES ⭐⭐
 
 ### Current: score=11.62, memory=643,378
 ### Target: score=12.5-13 (gain +0.9-1.4 pts)
 
 ---
 
-## BLOCKED:
-- ❌ Task 219 — fails hidden tests
-- ❌ Task 366 — filesize >1.44MB
-- ❌ Tasks needing CC detection (44, 76, 77, 101, 157, 182, 233)
+## ⭐ #5: TASK 023 — CONV7x7 TRAINING (needs Kaggle GPU)
+
+### Current: score=11.06 | Target: 12.5-13.2 (gain +1.4-2.1)
+### Status: Script ready, needs run with --hidden 128 --n-train 50000
+### Script: `medal-solvers/build_tasks/train_task023_final.py`
+### DO NOT attempt fp16 surgery or reduced iterations (both proven to fail)
+
+---
+
+## TOTAL POTENTIAL: +6.6 to +11.1 pts (tasks 1-4 alone = +5.6 to +9.1)
 
 ---
 
 ## RULES:
-- ❌ Don't use static scoring — off by 50%+. Profile on Kaggle only.
+- ❌ NEVER use static scoring — profile on Kaggle ONLY
 - ❌ Don't submit with <100% accuracy on known + generated examples
-- ❌ Don't train Conv weights expecting them to generalize without 5000+ examples
+- ❌ Don't work on task 219
 - ✅ Hand-craft weights using bias threshold trick where rule is understood
-- ✅ Use iterative small-kernel approach (like task 023's proven algorithm)
-- ✅ Verify on ALL examples + 5000+ random inputs
+- ✅ Use per-channel [1,1,H,W] to minimize memory
+- ✅ Use MaxPool propagation for iterative algorithms
+- ✅ Use delta approach for output (input + delta)
 - ✅ Profile on Kaggle before submitting
+- ✅ Verify on ALL examples + 5000+ random inputs