Rewrite docs: honest state, real plan (reverse-engineer 6043 into DSL solvers)

own-solver/README.md

@@ -1,65 +1,53 @@
-# NeuroGolf Solver v9.2
+# NeuroGolf Solver
 
 Builds minimal ONNX networks for ARC-AGI tasks. Modular Python package with opset 10-17, zero-cost Slice-based transforms + embedded trained weights.
 
-**Public LB: 5560.70** (V3 optimizer, 400/400 tasks submitted)
+**Public LB: 5560.70** (V3, 400/400 tasks)
 
 ## Target: LB > 6043
 
-Gap: +483 pts. Strategy: reduce intermediate tensor count (score = 25 - ln(memory + params)).
+submission-6043.zip already scored 6043 on Kaggle. It's in this repo.
+The models inside it are NOT generated by our solver — they come from an external source.
+To beat 6043 with OUR solver, we need to reverse-engineer those models into DSL solver code.
 
-## Scoring Formula (CONFIRMED — May 14 neurogolf_utils.py)
+## Scoring Formula (neurogolf_utils.py, May 14)
 
 ```python
 score = max(1.0, 25.0 - math.log(max(1.0, memory + params)))
 ```
-- memory = ORT Profiler measured tensor sizes (all intermediates, excludes input/output)
-- params = total initializer + Constant element count
-- **NO MACs** (removed May 4 after exploit)
+- memory = ORT Profiler intermediate tensor bytes (excludes input/output)
+- params = initializer element count + Constant element count
+- Initializer BYTES are NOT in memory — only element counts go to params
+- NO MACs (removed May 4)
 
-## Session Summary
+## What We Have
 
-**V90 (solver): LB 5322.64**
-- 385 tasks solved by solvers + 15 fallback reference models
-- All 400 tasks submitted
+| Source | LB Score | Origin |
+|--------|----------|--------|
+| submission-6043.zip | 6043 | External. Not from our solver. Needs reverse-engineering. |
+| submission-5743.zip | 5743 | External. Partially reverse-engineered into wave22. |
+| submission-v90.zip | 5322 | Our solver output. |
+| V3 (merged) | 5560 | merge_best_of_both.py + stage1_trivial_optimizer.py |
 
-**V3 (optimizer): LB 5560.70 (+238 pts)**
-- Merged with 5743 (80 tasks use compact 5743 models)
-- Stage 1 trivial optimization (6 tasks at 18-25 pts)
-- Proved: fewer nodes = less memory = higher score
-
-## Architecture
-
-| Component | Tasks | Description |
-|-----------|-------|-------------|
-| Wave 22 (automated) | 143 | Compact reverse-engineered from 5743 (15-50 nodes) |
-| Waves 17-21 (automated) | 78 | Learned weights + analytical graphs |
-| Analytical (waves 1-16) | 129 | Hand-coded + reverse-engineered DSL |
-| Conv lstsq | 35 | Least-squares fitted kernels |
-| Fallback reference | 15 | Reference models for crash-tasks |
-
-## Optimization Tools
-
-| Tool | Purpose | Impact |
-|------|---------|--------|
-| `merge_best_of_both.py` | Pick compact model from 5743 vs V90 | +200 pts |
-| `stage1_trivial_optimizer.py` | Rebuild simple tasks (transpose=25pts) | +38 pts |
-| `sign_corrected_conv.py` | Direct Conv→output for conv_var | +8 pts |
-| `fp16_optimizer.py` | fp16 weight conversion | needs Cast fix |
-
-## Usage
+## Pipeline (current, working)
 
 ```bash
-# Standard solver run:
-python -m neurogolf_solver.main --device cpu --conv_budget 5 \
-  --data_dir $DATA_DIR --kaggle --output_dir submission --fallback_zip fallback_unsolved.zip
-
-# Post-processing optimization:
-python sign_corrected_conv.py --data_dir $DATA_DIR --output_dir submission
-python stage1_trivial_optimizer.py --input_zip submission.zip --data_dir $DATA_DIR --output_zip out.zip
-python merge_best_of_both.py --sub_5743 5743.zip --sub_v90 out.zip --data_dir $DATA_DIR --output_zip submission.zip
+python merge_best_of_both.py \
+  --sub_5743 submission-5743.zip \
+  --sub_v90 submission-v90.zip \
+  --data_dir /kaggle/input/competitions/neurogolf-2026 \
+  --output_zip /kaggle/working/merged.zip
+
+python stage1_trivial_optimizer.py \
+  --input_zip /kaggle/working/merged.zip \
+  --data_dir /kaggle/input/competitions/neurogolf-2026 \
+  --output_zip /kaggle/working/submission.zip
 ```
 
+## What Needs to Happen
+
+Reverse-engineer submission-6043.zip models into solver code. See NEXT_AGENT.md.
+
 ## Repo
 
 https://huggingface.co/rogermt/neurogolf-solver