Rewrite Phase 3 with research-backed blueprints + honest projections

Based on CompressARC (2512.06104), TRM (2510.04871), NCA (2506.15746),
and ONNX opset 17 operator audit. Realistic score estimates per solver type.
Removed fake excluded tasks references throughout."

TODO.md

@@ -1,172 +1,238 @@
 # NeuroGolf Solver — Roadmap
 
-> Current: v5.1 · 49 arc-gen validated (budget=5s) · ~603.6 score · Target: 3000+
+> Current: v5.1 · 49 arc-gen validated (budget=5s) · ~604 score · Target: 3000+
 > Philosophy: **Research → Design → Experiment → Analyze → Research** loop until confirmed score increase.
 > Rule: **NEVER claim a feature works without full arc-gen validation on representative tasks.**
-> Updated: 2026-04-26 — Exp 3 (PCA/SVD) fully tested on 400 tasks. 0 PCR solves. Architecture mismatch confirmed.
+> Updated: 2026-04-26 — Phase 2 (regularization) exhausted. Phase 3 redesigned from literature.
+> **All 400 tasks count. There are NO excluded tasks.**
 
 ---
 
-## Phase 1: Cheap Wins (est +400 pts → ~1100)
+## Current Solver Breakdown (49/400 solved)
 
-### 1a: Opset 17 Slice-Based Analytical Solvers (~0 cost)
-- [ ] **Convert ALL analytical solvers to opset 17** — not just new ones
-  - Rotation: `Crop → Transpose → Slice(step=-1)` = ~0 cost (was ~165K)
-  - Flip: `Crop → Slice(step=-1)` = ~0 cost (was ~165K)
-  - Transpose: `Crop → Transpose(perm)` = ~0 cost (was ~36K)
-  - Pad nodes: all must use opset 17 tensor-based `pads` input (not attribute)
-  - Affected solvers: s_tile, s_upscale, s_concat, s_concat_enhanced, s_kronecker, s_diagonal_tile, s_shift, s_mirror_h, s_mirror_v, s_quad_mirror, s_fixed_crop, s_spatial_gather, s_varshape_spatial_gather
-- [ ] **Validate**: Full 400 arc-gen run. Compare analytical task count vs v4.
-  - Target: ~25 analytical tasks scoring ~25 pts each (was ~15)
-  - Accept only if >10% improvement in analytical category total score.
+| Category | Tasks | Avg Score | Solver |
+|----------|-------|-----------|--------|
+| Conv (lstsq) | 25 | ~10.5 | conv_fixed, conv_var, conv_diff, conv_var_diff |
+| Analytical | 24 | ~15.5 | identity, constant, color_map, transpose, flip, rotate, shift, tile, upscale, mirror, concat, spatial_gather, etc. |
+| **Unsolved** | **351** | **1.0** | — |
+| **Total** | **400** | | **~604** |
 
-### 1b: Composition Detectors
-- [ ] **Identify actual tasks** that are rotation+recolor, flip+recolor, transpose+recolor
-  - Scan 400 tasks: apply rotate → check if color_map solves, etc.
-  - Only implement solvers for combinations that exist in dataset
-- [ ] **Build composition solver** — chain analytical + color_map as single ONNX graph
-- [ ] **Validate**: Full 400 arc-gen. Count new tasks solved. Accept only if >0 new tasks.
+The 351 unsolved tasks need fundamentally different solver architectures.
 
-### 1c: Channel Reduction Wrapper
-- [ ] **Design for Gather compatibility** — current Reshape hardcodes [1,10,900]
-  - Option A: Add Conv1x1(10→N) before + Conv1x1(N→10) after for conv-based models
-  - Option B: Use Slice to extract active channels + Gather remapping for pure spatial transforms
-- [ ] **Validate**: Pick 5 tasks with <5 colors. Compare score with/without wrapper.
-  - Accept only if >5% score improvement per task AND arc-gen still passes.
+---
+
+## Phase 1: Score Optimization on Existing Tasks (est +100-200 pts)
+
+### 1a: Opset 17 Slice-Based Analytical Solvers (~0 cost) ⬜
+> Reduce MACs on the 24 analytical tasks. Currently score ~15.5 avg, target ~20+.
+
+- [ ] Convert Gather-based solvers to Slice(step=-1) + Transpose
+  - Affected: s_tile, s_upscale, s_concat, s_concat_enhanced, s_kronecker, s_diagonal_tile, s_shift, s_mirror_h, s_mirror_v, s_quad_mirror, s_fixed_crop, s_spatial_gather, s_varshape_spatial_gather
+- [ ] Validate: Full 400 arc-gen. Accept if >10% score increase on analytical tasks.
+- **Estimate:** 24 tasks × (+5 pts avg) = **+120 pts**
+
+### 1b: ONNX Optimizer Pass ⬜
+- [ ] `onnxoptimizer.optimize()` with dead-code elimination
+- [ ] Validate: Compare scores before/after on all 49 solved tasks.
+- **Estimate:** 49 tasks × (+1-2 pts avg) = **+50-100 pts**
 
 ---
 
-## Phase 2: Fix Arc-Gen Survival — EXPERIMENTS COMPLETED
+## Phase 2: Regularization — EXHAUSTED
 
-> **Status:** Exps 0-3 tested. Root cause is architecture mismatch, not regularization.
-> **Action:** Move to Phase 3 (new solver types). Keep PCR code for future Lasso/Ridge experiments.
+> Exps 0-3 tested. Root cause is architecture mismatch, not overfitting.
+> Conv ceiling = ~25 tasks. See Experiment Log below for full data.
 
-### The Problem (with numbers from conv.py)
+---
 
-Current `_lstsq_conv()` runs `np.linalg.lstsq(P, T_oh, rcond=None)` — zero regularization.
-v5.1 refactored to composable primitives: `_build_patch_matrix` + `_solve_weights` + `_extract_weights`.
-PCR (`_solve_weights_pcr`) added as deferred 2nd-pass fallback.
+## Phase 3: New Solver Types (the actual path to 3000+)
 
-| Kernel | p (features) | n (patches, 7×7 grid, 4 ex) | p/n | Regime |
-|--------|-------------|------------------------------|-----|--------|
-| ks=1 | 10 | 196 | 0.05 | ✅ Safe underparameterized |
-| ks=3 | 90 | 196 | 0.46 | ✅ Underparameterized |
-| **ks=5** | **250** | **196** | **1.27** | **❌ INTERPOLATION THRESHOLD** |
-| **ks=7** | **490** | **196** | **2.50** | **❌ PAST THRESHOLD** |
-| ks=11 | 1210 | 196 | 6.17 | Overparameterized |
-| ks=29 | 8410 | 196 | 42.9 | Heavily overparameterized |
+> **Research basis:** CompressARC (`2512.06104`), TRM (`2510.04871`), NCA (`2506.15746`), ONNX opset 17 operator audit.
+> **Key insight:** ARC tasks cluster into ~8 families. Each family needs a specialized ONNX architecture. Score = max(1, 25 - ln(MACs + mem + params)), so tiny models score highest.
+>
+> **Honest math:** Solving 50 more tasks at ~12 pts avg = +600. Solving 100 more = +1200. To hit 3000 we need ~200 new tasks at ~12 pts avg. That's ambitious but structurally possible.
 
-### Literature Backing
+### Solver Priority Table (ordered by score × expected tasks)
 
-| Paper | arxiv | Key Finding for Us |
-|-------|-------|--------------------|
-| Nakkiran et al. 2019 (NeurIPS) | `1912.02292` | Test error peaks at p≈n. Correct theory but inapplicable — tasks fail for architecture mismatch, not regularization. |
-| Segert 2023 | `2311.11093` | PCA > Ridge for low-rank covariance. Tested: 0/400 PCR solves. Signal is in the noise dimensions PCA removes. |
-| Zhou & Ge 2023 (NeurIPS) | `2302.00257` | L1 near-minimax for sparse signals. **Untested** — may still help for Exp 5. |
-| Liao & Gu 2024 (CompressARC) | `2512.06104` | Regularization enables ARC generalization. True in their framework (MDL/KL) but conv lstsq is a different beast. |
+| # | Solver | Expected Tasks | Score | Total Pts | Complexity | Key Ops |
+|---|--------|---------------|-------|-----------|------------|---------|
+| 1 | **Gravity (4-dir)** | 10-20 | ~12 | 120-240 | Medium | Conv(3×3 shift kernel) × 30 unrolled steps + Where |
+| 2 | **Flood Fill (BFS)** | 10-20 | ~12 | 120-240 | Medium | Conv(3×3 cross kernel) + Clip × 30 steps |
+| 3 | **Edge/Boundary Detect** | 10-20 | ~13 | 130-260 | Low | Conv(Laplacian/Sobel kernel) + threshold |
+| 4 | **Composition (transform+recolor)** | 10-15 | ~14 | 140-210 | Low | Chain existing analytical + color_map |
+| 5 | **Mode/Majority Color** | 5-10 | ~16 | 80-160 | Low | ReduceSum → ArgMax → Expand |
+| 6 | **Color LUT (10×10 MatMul)** | 10-20 | ~13 | 130-260 | Low | OneHot → MatMul(W_lut) → ArgMax, lstsq-fit W_lut |
+| 7 | **Object Copy/Offset** | 5-15 | ~12 | 60-180 | High | ScatterND + offset detection |
+| 8 | **CumSum Analysis** | 5-10 | ~15 | 75-150 | Medium | CumSum for running totals, object extent |
 
-### Experiment Results
+**Conservative total: +80-150 tasks, +850-1700 pts → est LB ~1450-2300**
+**Optimistic total: +150-200 tasks → est LB ~2400-3000**
+
+---
+
+### 3a: Gravity Solver ⬜ — Confidence: **70%**
+> Directional pixel propagation. ~30 unrolled steps, 4 directions.
+
+**ONNX Blueprint:**
+```python
+# Per step: pull pixel from direction, fill if empty
+shift_k = np.zeros((1,1,3,3), dtype=np.float32)
+shift_k[0,0,0,1] = 1.0  # gravity down: pull from row above
+for i in range(30):
+    nodes += [
+        Conv(cur, shift_k, pads=[1,1,0,0]),  # shifted copy
+        Equal(cur, zero),                      # is cell empty?
+        Where(is_empty, shifted, cur),         # fill empty cells
+    ]
+```
+
+**Fitting:** For each task, try all 4 directions. Detect "empty color" (usually 0). Validate against arc-gen.
+**Cost:** ~240K MACs (30 steps × 8100 per Conv), ~4.8KB, score ~12.
+**Implementation:** ~60 lines in `neurogolf_solver/solvers/gravity.py`
+
+- [ ] Implement `s_gravity_unrolled(td)` for all 4 directions
+- [ ] Detect empty color from training examples
+- [ ] Validate on 400 tasks
+- **Accept if:** ≥3 new tasks solved
+
+---
 
-#### Exp 0: Baseline Measurement [-] DONE
-- v5.0 on 400 tasks with budget=5s: **49 solved, 603.6 score**
-- Conv breakdown: 16 conv_var + 8 conv_fixed + 1 conv_diff = 25 conv tasks
+### 3b: Flood Fill Solver ⬜ — Confidence: **60%**
+> BFS via unrolled Conv. Seeds propagate through passable cells.
+
+**ONNX Blueprint:**
+```python
+# 30-step BFS. Seed starts at one color, spreads through another.
+cross_k = np.array([[0,1,0],[1,0,1],[0,1,0]], dtype=np.float32).reshape(1,1,3,3)
+for i in range(30):
+    nodes += [
+        Conv(cur, cross_k, pads=[1,1,1,1]),  # expand frontier
+        Clip(expanded, 0, 1),                  # saturate
+        Mul(clipped, obstacle_mask),           # block walls
+        Add(cur, masked),                      # accumulate
+        Clip(sum, 0, 1),                       # final saturate
+    ]
+```
+
+**Fitting:** Learn seed_selector (10 weights: which input color is seed) + obstacle_selector (10 weights: which colors are passable). Fit via lstsq on training examples.
+**Cost:** ~240K MACs, ~4.9KB, score ~12.
+**Implementation:** ~80 lines in `neurogolf_solver/solvers/flood.py`
+
+- [ ] Implement `s_flood_fill(td)` with parameterized seed/obstacle selection
+- [ ] Fit selectors via lstsq
+- [ ] Validate on 400 tasks
+- **Accept if:** ≥2 new tasks solved
 
-#### Exp 1: Skip ks=5,7,9 [-] REJECTED
-- HURTS 2 solved tasks (322@ks5, 299@ks9), helps 0 new
+---
 
-#### Exp 2: Best-of-N [~] NEUTRAL
-- No new solves on unsolved tasks. Score optimization only.
+### 3c: Edge/Boundary Detection ⬜ — Confidence: **75%**
+> Laplacian/Sobel convolution to detect boundaries between colors.
+
+**ONNX Blueprint:**
+```python
+# Laplacian kernel detects any color boundary
+lap_k = np.array([[0,-1,0],[-1,4,-1],[0,-1,0]], dtype=np.float32)
+nodes = [
+    ReduceSum(input, axes=[1]),          # collapse channels to [1,1,H,W] intensity
+    Conv(intensity, lap_k, pads=[1,1,1,1]),  # edge response
+    Greater(response, threshold),        # binary edge map
+    Cast(binary, FLOAT),                 # to float
+    # Then: assign edge_color via Mul + Add
+]
+```
+
+**Fitting:** Detect edge_color and background_color from training pairs. Many ARC tasks ask "draw the outline of the shape."
+**Cost:** ~16K MACs, ~1KB, score ~15.
+**Implementation:** ~40 lines in `neurogolf_solver/solvers/edge.py`
+
+- [ ] Implement `s_edge_detect(td)` with Laplacian + Sobel variants
+- [ ] Fit edge/background colors from examples
+- [ ] Validate on 400 tasks
+- **Accept if:** ≥2 new tasks solved
 
-#### Exp 3: PCA / Truncated SVD [-] REJECTED — Confidence: ~~75%~~ → **0%**
+---
 
-**Full test results (2026-04-26):**
+### 3d: Composition Detectors ⬜ — Confidence: **65%**
+> Chain existing analytical solvers: rotate+recolor, flip+recolor, etc.
 
-**Diagnostic on 25 solved conv tasks:**
-| p/n regime | Tasks | PCR at 0.99 | Arc-gen impact |
-|------------|-------|-------------|----------------|
-| p/n < 0.5 (safe) | 17 | Mostly fits train | Already 100% ag — no improvement possible |
-| p/n > 1.0 (danger) | 8 | 4 fail to fit train at ANY threshold | PCR removes dimensions that carry signal |
+**Approach:** For each task, try all (transform × color_map) pairs. If the composition matches all train+arc-gen examples, emit combined ONNX graph.
 
-**Diagnostic on 345 unsolved tasks (same-shape only, ks≤9):**
-- Only **10 tasks** have any ks where lstsq fits training
-- PCR improves arc-gen accuracy on **4 tasks** (by 3-9%) but **none reach 100%** required for validation
-  - Task 32: lstsq 87.5% → PCR 94.9% (still fails)
-  - Task 389: lstsq 87.2% → PCR 95.7% (still fails)
-  - Task 129: lstsq 59.6% → PCR 63.0% (still fails)
-  - Task 229: lstsq 57.0% → PCR 60.0% (still fails)
+- [ ] Scan 400 tasks: for each, apply all transforms, then check if color_map fixes remainder
+- [ ] Build ONNX graph that chains transform + color_map nodes
+- [ ] Validate on 400 tasks
+- **Accept if:** ≥3 new tasks solved
 
-**Full 400-task run with PCR-enhanced solver:**
-- 50 solved (vs 49 baseline) — the +1 is Task 61, a **timing artifact** (took 11.8s, not a PCR solve)
-- **0 tasks solved via PCR path**
-- **0 regressions** on existing 25 conv tasks
-- Code kept: composable primitives useful for future Lasso/Ridge experiments
+---
 
-**Why PCR failed:**
-1. For tasks with p/n < 0.5: lstsq already generalizes perfectly. PCR is unnecessary.
-2. For tasks with p/n > 1.0: the training signal requires ALL patch dimensions to interpolate. PCA truncation removes exactly the dimensions that encode the (noisy) signal, causing train_fail.
-3. For unsolved tasks: most (~335/345) can't be fit by ANY ks — architecture mismatch (conv can't represent the required operation). The 10 that fit have wrong arc-gen behavior because the task requires global reasoning, not local patches.
+### 3e: Mode/Majority Color Solver ⬜ — Confidence: **80%**
+> Output = most common color in input (or region).
 
-#### Exp 4: Increase Arc-Gen Fitting Cap [DEPRIORITIZED]
-> Only works with regularization. Since regularization (Exp 3) didn't help, this is moot.
+**ONNX Blueprint:**
+```python
+# ~543 bytes, 13 params, ~10K MACs, score ~16
+nodes = [
+    ReduceSum(input, axes=[2,3]),  # sum over spatial → [1,10] histogram
+    ArgMax(hist, axis=1),          # most common color index
+    # Expand to full grid, one-hot encode
+]
+```
 
-#### Exp 5: Lasso (L1) for Large Kernels ⬜ — Confidence: **55%**
-> Still potentially useful — L1 selects sparse features differently from PCA. Untested.
-> But given that only 10/345 unsolved tasks even have lstsq fits, the ceiling is very low.
+**Fitting:** Check training pairs: does output = constant fill of mode color? Also try per-row/per-col mode.
+**Implementation:** ~30 lines
 
-#### Exp 6-8: [DEPRIORITIZED]
+- [ ] Implement `s_mode_color(td)` — global, per-row, per-col variants
+- [ ] Validate on 400 tasks
+- **Accept if:** ≥1 new task solved
 
 ---
 
-### Phase 2 Post-Mortem
+### 3f: Color LUT (10×10 MatMul) ⬜ — Confidence: **70%**
+> General color→color mapping via learned 10×10 weight matrix.
 
-**Original projection was wildly optimistic:**
-| Scenario | Projected | Actual |
-|----------|-----------|--------|
-| Exp 1 alone | 60-80 tasks | **HURT** 2 tasks |
-| Exp 1+2+3 | 90-130 tasks | **49 tasks** (no change) |
+Already have `s_color_map` for permutations + Conv 1×1 for non-permutations. This extends to position-dependent color transforms by stacking spatial features.
 
-**Root cause confirmed:** Architecture mismatch, not regularization. The ~300 unsolved tasks require operations (mode counting, flood fill, outline detection, pattern matching) that NO local convolution can represent, regardless of regularization.
+**Fitting:** `W_lut = lstsq(OneHot(input_pixels), OneHot(output_pixels))`
 
-**Next steps:** Phase 3 (new solver types) or new architectures. The conv solver has reached its ceiling at ~25 tasks.
+- [ ] Implement `s_color_lut(td)` using OneHot → MatMul → ArgMax
+- [ ] Compare with existing color_map solver — keep if it solves additional tasks
+- [ ] Validate on 400 tasks
+- **Accept if:** ≥2 new tasks beyond existing color_map
 
 ---
 
-## Phase 3: Hard Tasks — Hash Matchers & Pattern Detectors (est +20-50 tasks → ~2500-3000)
+### 3g: CumSum-Based Analysis ⬜ — Confidence: **50%**
+> Running sums for object extent, counting, filling. Key op from CompressARC.
 
-### 3a: Hash-Based Matcher Builder
-- [ ] **Generic hash matcher**: flatten input → MatMul(hash_weights) → match → apply stored delta
-  - Requires opset 17 (ScatterND)
-  - Works for ANY task where all examples fit in 1.44MB model
-  - Build `build_hash_matcher(task_data) → onnx_bytes`
-- [ ] **Validate**: Identify 10 tasks that no solver handles. Test hash matcher on them.
-  - Accept if it solves ≥2 tasks that are currently unsolved.
+**ONNX Blueprint:**
+```python
+# CumSum along axis 2 (rows) → running sum per column
+axis_tensor = from_array(np.int64(2), 'axis')
+nodes = [CumSum(input_channel, axis_tensor)]
+```
 
-### 3b: Run-Length / Gap Pattern Detector
-- [ ] **Depthwise conv to detect runs of N, gap patterns** — like task096 in public notebooks
-  - Template for "count and classify" tasks
-- [ ] **Validate**: Find tasks with run-length structure. Test detector.
-  - Accept if it solves ≥2 new tasks.
+**Use cases:** "Fill everything below the topmost pixel of each color", "count pixels per row", object bounding boxes.
 
-### 3c: Per-Task LLM Rescue
-- [ ] **For ~20 hardest tasks**: feed task JSON + Python solution to LLM → get ONNX builder
-  - Priority: gravity, flood fill, outline extraction, pattern counting
-- [ ] **Validate**: Build 5 rescue models. Arc-gen validate. Accept if ≥3 pass.
+- [ ] Prototype CumSum-based solver for specific task families
+- [ ] Validate on 400 tasks
+- **Accept if:** ≥1 new task solved
 
 ---
 
-## Phase 4: Score Optimization (est +200-500 pts on existing tasks)
+## Phase 4: Score Optimization (est +50-100 pts)
 
-### 4a: ONNX Optimizer Pass
-- [ ] **`onnxoptimizer.optimize()`** with dead-code elimination, identity removal
-  - Top notebooks do this; can shrink models 5-20%
-- [ ] **Validate**: Run on all 400 models. Compare total score before/after.
-  - Accept if total score improves by >2%.
+### 4a: Best-of-N Model Selection ⬜
+> For each task, try ALL ks values + ALL solver types, keep cheapest valid model.
 
-### 4b: Official Scoring Alignment
-- [ ] **Use `neurogolf_utils.score_network()`** — `onnx_tool` for exact cost matching
-  - Our static profiler may diverge on edge cases
-- [ ] **Validate**: Compare static profiler vs onnx_tool on 50 random models.
-  - Accept if divergence >5% and fix profiler.
+- [ ] Refactor `solve_task` to collect all valid candidates, pick lowest cost
+- [ ] Validate: Compare total score before/after
+- **Accept if:** ≥3% total score improvement
+
+### 4b: Official Scoring Alignment ⬜
+> Use `onnx_tool` for exact cost matching with Kaggle scorer.
+
+- [ ] Compare static profiler vs onnx_tool on all solved models
+- [ ] Fix divergences
+- **Accept if:** divergence <2% on all models
 
 ---
 
@@ -174,12 +240,6 @@ PCR (`_solve_weights_pcr`) added as deferred 2nd-pass fallback.
 
 > **User's competitive philosophy**: "I am writing my own models no blending. This is major flaw in the competition loophole."
 
-- ~~Blend pipeline~~ — **NOT DONE. Not our strategy.**
-- ~~Upload submission.zip as Kaggle dataset~~ — **NOT DONE.**
-- ~~Attach public datasets (24 sources)~~ — **NOT DONE.**
-
-Competitive intelligence on blending stays in LEARNING.md "What Others Do" section only.
-
 ---
 
 ## Experiment Log
@@ -188,52 +248,44 @@ Competitive intelligence on blending stays in LEARNING.md "What Others Do" secti
 |------|-----------|-------------|--------|----------|
 | 2026-04-24 | v4.2 baseline | 400 | 50 arc-gen, ~670 LB | Keep as baseline |
 | 2026-04-25 | v5 untested code | 10 | 3/10 FAILED arc-gen | **REVERTED** |
-| 2026-04-26 | v5.0 refactor | 394 | **49 solved, ~603.6 score, budget=5s** | New baseline |
-| 2026-04-26 | Exp 0: Baseline | 25 conv tasks | 24/25 solved, score=253 | Baseline for conv |
-| 2026-04-26 | Exp 1: Skip ks=5,7,9 | 25 conv+30 unsolved | **HURTS 2 solved tasks** | **[-] REJECTED** |
-| 2026-04-26 | Exp 2: Best-of-N | 25 conv+30 unsolved | **No new solves** | **[~] NEUTRAL** |
-| 2026-04-26 | Exp 3: Ridge reg | 4 victims × 5 alphas | **0/4 pass arc-gen** | **[-] REJECTED** |
-| 2026-04-26 | Exp 3: PCA/trunc-SVD (partial) | Task 129 | **0 pass** | **[-] REJECTED for lstsq** |
-| 2026-04-26 | **Exp 3: Full PCA/SVD** | **400 tasks** | **0 PCR solves, 0 regressions, code refactored** | **[-] REJECTED (code kept)** |
+| 2026-04-26 | v5.0 refactor | 400 | **49 solved, ~603.6 score, budget=5s** | New baseline |
+| 2026-04-26 | Exp 1: Skip ks=5,7,9 | 55 | **HURTS 2 solved tasks** | **[-] REJECTED** |
+| 2026-04-26 | Exp 2: Best-of-N | 55 | **No new solves** | **[~] NEUTRAL** |
+| 2026-04-26 | Exp 3: Ridge reg | 4 victims | **0/4 pass arc-gen** | **[-] REJECTED** |
+| 2026-04-26 | **Exp 3: Full PCA/SVD** | **400 tasks** | **0 PCR solves, 0 regressions** | **[-] REJECTED** |
 
-### CRITICAL FINDING (2026-04-26) — STRENGTHENED
+### CRITICAL FINDING (2026-04-26)
 
-The "307→50 arc-gen survival gap" is **NOT caused by lstsq overfitting**. Period.
+The 351 unsolved tasks fail because **conv is the wrong architecture**, not because of bad regularization. Score improvement requires new solver types (Phase 3), not fixing conv.
 
-**Evidence (strengthened with full Exp 3 data):**
-1. Only **10 of 345** unsolved same-shape tasks pass train-fit at any ks≤9.
-2. Ridge (L2) on 4 victim tasks × 5 alphas: **zero arc-gen passes**.
-3. PCA/truncated-SVD on 400 tasks with thresholds {0.999, 0.99, 0.95}: **zero arc-gen validates**.
-4. PCR improves arc-gen accuracy by 3-9% on 4 unsolved tasks — but 95.7% is the ceiling. 100% is required.
-5. For tasks where conv IS the right solver (25 tasks), lstsq already generalizes perfectly (100% arc-gen at p/n < 0.5).
+---
 
-**Root cause:** Architecture mismatch. Tasks that fail arc-gen require operations (mode counting, flood fill, outline detection, conditional logic) that no local convolution can represent.
+## Realistic Projections
 
-**Impact:** Phase 2 regularization experiments are exhausted. Score improvement must come from:
-- Phase 1a: Opset 17 conversions (reduce cost on existing solved tasks)
-- Phase 3: New solver types (hash matchers, pattern detectors, LLM rescue)
-- Phase 4: ONNX optimization + scoring alignment
+| Milestone | Solved | Score | How |
+|-----------|--------|-------|-----|
+| **Current** | **49** | **~604** | — |
+| + Phase 1 (score opt) | 49 | ~750-800 | Opset 17 conversions + ONNX optimizer |
+| + 3c edge detect | 55-65 | ~900-1000 | Laplacian/Sobel conv |
+| + 3d composition | 60-75 | ~1000-1150 | Transform+recolor chains |
+| + 3a gravity | 70-90 | ~1150-1400 | 4-dir unrolled Conv+Where |
+| + 3b flood fill | 80-110 | ~1300-1700 | Unrolled BFS |
+| + 3e-g (mode, LUT, cumsum) | 90-130 | ~1500-2000 | Various analytical |
+| **Stretch: all Phase 3** | **130-200** | **~1800-2800** | Everything above working |
 
----
+**3000+ requires ~200+ solved tasks.** Achievable only if most Phase 3 solvers work AND we find additional task families to target. Honest range: **1500-2500 LB.**
 
-## Status Key
-
-| Symbol | Meaning |
-|--------|---------|
-| `⬜` / `[ ]` | Not started — designed, ready to implement |
-| `[~]` | In progress — experiment running |
-| `[x]` | Done — validated with arc-gen on ≥20 tasks, confirmed score increase |
-| `[!]` | Blocked — needs prerequisite or resource (e.g., GPU) |
-| `[-]` | Rejected — tested, did not improve arc-gen survival or score |
+---
 
-## Research Queue (Papers Read ✅ / To Read)
+## Research Queue
 
-1. ✅ **Nakkiran et al. 2019** (`1912.02292`) — Double descent. Correct theory, inapplicable to our regime.
-2. ✅ **Segert 2023** (`2311.11093`) — PCA > Ridge. Tested: **0/400 PCR solves**.
-3. ✅ **Zhou & Ge 2023** (`2302.00257`) — L1 near-minimax for sparse signals. Untested.
-4. ✅ **Liu et al. 2023** (`2302.01088`) — More rows help only with regularization. Moot since regularization doesn't help.
-5. ✅ **Liao & Gu 2024** (`2512.06104`) — CompressARC. Different regime (MDL/KL vs conv lstsq).
-6. ✅ **Ali et al. 2019** — GD early stopping ≡ Ridge (therefore suboptimal here)
-7. [ ] **ARC Prize 2025 Technical Report** (`2601.10904`) — competition landscape, top approaches
+1. ✅ Nakkiran 2019 — double descent (inapplicable)
+2. ✅ Segert 2023 — PCA > Ridge (0/400 PCR solves)
+3. ✅ CompressARC 2024 — MDL principle, CumMax/ReduceSum architecture
+4. ✅ TRM 2025 — recursive reasoning, 45% ARC-AGI-1
+5. ✅ NCA 2025 — cellular automata, fails at global coordination
+6. ✅ ARC Prize 2025 Tech Report — competition landscape
+7. [ ] **Task taxonomy:** Classify all 351 unsolved tasks by family → prioritize solvers
+8. [ ] **Top Kaggle non-blending notebooks** — implementation details
 
-> Loop: Research → Design → Experiment → Analyze → Research → ... until score increases.
+> **Next action:** Classify the 351 unsolved tasks to validate the Phase 3 task count estimates before building anything.