docs/REVISION_LOG.md

# Geoderma Revision Log — GEODER-D-26-01032

**Branch:** `bestrun-bands`
**Manuscript:** *Spatiotemporal Gated Transformer for High-Resolution Soil Organic Carbon Mapping Using Multi-Temporal Remote Sensing*
**Editor:** Budiman Minasny
**Decision:** Major Revision
**Resubmission deadline:** 2026-05-29

This document is a phase-by-phase log of every code change, methodological
decision, finding, and strategic discussion in the revision arc. Commit
SHAs are referenced where applicable. Phases are roughly chronological but
ordered by theme.

---

## 0. Starting state

Before this revision arc, `bestrun-bands` held:

- `8e59df2 SGT V2`: per-band input projection, wider head with linear skip
- `0acdb60`: module-level distributed init guarded for importable use
- `0db68f3 kfold` rewrite: latitude-decile spatial folds, train.py recipe
- `26eb6b6`: deterministic seed before split/rebalance
- `e706287`: NUM_LAYERS=3, NUM_HEADS=4 set as base architecture
- `dc493b8`: 20-channel band expansion ported across 9 sibling configs
  (2DCNN, 3DCNN, Baselines, CNNLSTM, SimpleTransformer, TFT, VAE,
  FoundationalModels, balancedDataset) — **but SGT's own config.py was
  missed in this commit** (caught and fixed later in Phase 2).

The k-fold pipeline existed but ran sequentially: one fold at a time per
job, ten folds × ~per-fold time. No production-mapping pipeline yet,
no architecture comparison, no spatial-CV bug discovery yet.

---

## 1. Parallel orchestrator (`feb2a8d`)

**Problem.** The k-fold orchestrator capped concurrency at one fold per
GPU. With 4 GH200s and 10 folds, three sequential waves (4+4+2). Per-fold
SGT V2 uses ~1.5 GB of HBM; the 96 GB GH200s were ~95% idle.

**Decision.** User chose 3 folds/GPU (12 slots, 10 used) and **no MPS**
(rely on CUDA's default time-slicing — sufficient for a ~1.2M-param
model where memory and bandwidth aren't the bottleneck).

**Change in `rebuttal/gpu_experiments/spatial_kfold/run_folds_parallel.py`:**
- New `--folds-per-gpu` flag, default 1 for backward compatibility.
- Slot-keyed bookkeeping (`slot_idx → gpu_id` via `slot_idx % num_gpus`)
  replacing the previous `gpu_id`-keyed dicts.
- Each fold subprocess still pinned to exactly one GPU via
  `CUDA_VISIBLE_DEVICES=<one-id>`.

**Result.** Single command launches all 10 folds simultaneously:

```bash
python rebuttal/gpu_experiments/spatial_kfold/run_folds_parallel.py \
    --num-folds 10 --num-parallel 10 --folds-per-gpu 3 \
    -- <run_kfold.py args>
```

Wall time drops from `3 × per_fold_time` to `~1 × per_fold_time`. ~10–15
min total at 300 epochs, ~10 min at 30 epochs (which became the standard
screening epoch budget after later diagnostic work — see Phase 5).

---

## 2. Cherry-picked methodology pivot — sampling, cap, augmentation, stratified reporting (`fb58def`)

**Problem.** User flagged that the original methodology had cherry-picked
parameters: MAX_OC=150 (the dataset max — no soil-science justification),
`rebalance_min_ratio=0.75` with 128 qcut bins (heavy oversampling of
rare-tail rows by literal duplication). A reviewer-proof methodology was
needed.

**Decisions taken in this phase:**

### 2a. MAX_OC: 150 → 120 (WRB histosol threshold)

Reframing the cap from "the dataset max" to "the WRB international
threshold separating mineral from organic soils." Drops 154 rows (0.93%
of 16,514). Single sentence of soil-science justification.

`SpatiotemporalGatedTransformer/config.py:27`:
```python
MAX_OC = 120  # WRB histosol threshold (organic soils excluded).
```

### 2b. Replace qcut-rebalance with KDE-density sampler

`rebalance_by_oc_bin` was qcut-ing OC into 128 quantile bins and
upsampling rare bins by literal row duplication (memorization risk on a
heavy-tailed distribution).

Replacement: `WeightedRandomSampler` with weights
`w_i ∝ kde(log(OC_i))^(-α)`, normalized to mean 1. Yang et al. ICML 2021
("DenseLoss / DenseWeight"). KDE bandwidth = Scott's rule (default;
**not tuned** — important for defensibility). Default α = 0.5 = sqrt-
frequency weighting.

New CLI in `run_kfold.py`:
- `--sampler-mode {kde, qcut}` (default kde)
- `--alpha-density` (default 0.5)
- `--rebalance-min-ratio` retained for legacy qcut mode

### 2c. D4 spatial augmentation

`_AugmentingWrapper` applies a uniformly-sampled element of the dihedral
group D4 (4 rotations × 2 flips) on the (H, W) plane of each train-loader
sample. Test loader unchanged. Soil patches have no orientation prior;
this is label-preserving.

### 2d. Stratified-by-OC-band reporting

Pooled predictions across all 10 fold test sets, binned by actual OC into
`[0, 20)`, `[20, 40)`, `[40, 80)`, `[80, ∞)`, plus `all`. Added to
`kfold_results.md` and `kfold_results_summary.json`. Pre-empts the
reviewer question "where does the R² come from?" — answers regime-by-regime.

### 2e. Per-fold predictions saved to disk + `--aggregate-only` mode

Each fold subprocess now writes `fold_<i>_predictions.parquet`. The
orchestrator (after Phase 1) launches per-fold subprocesses in parallel;
without per-fold persistence, the cross-fold aggregation couldn't run.
New `--aggregate-only` mode reads every per-fold parquet and writes the
cross-fold summary.

Orchestrator (`run_folds_parallel.py`) now invokes
`python run_kfold.py --aggregate-only --num-folds 10 …`
automatically after all per-fold subprocesses complete.

---

## 3. Verification, GPU monitoring, first symptom diagnosis

After the methodology changes landed, the user submitted the full
parallel sweep at MAX_OC=120 with the new sampler. Initial concerns:

### 3a. Was the 20-band port complete?

User asked: *"is it using all 20 channels?"*

Investigation revealed `dc493b8` ("Port band expansion (20 channels) onto
bestrun") updated 9 sibling configs but **missed
`SpatiotemporalGatedTransformer/config.py`**. The SGT was training on the
old 6-band stack. Fixed in `25ae836`:
- `bands_list_order` extended to 20 entries (originals first, 14 appended)
- `SamplesCoordinates_Yearly` and `DataYearly` extended to length 20
- 14 new `*TensorDataYearly` and `*BandMatrixCoordinates_Yearly` path
  variables added

Verified via param count: SGT V2 at 6 channels = 1,228,627 params; at
20 channels = 1,234,115 params. (The grouped 1×1 input conv only adds
~5k params per +14 bands — empirical confirmation of the
band-extensibility claim that became Phase 8's headline.)

### 3b. GPU monitoring on the Slurm cluster

User asked how to watch GPU usage. Suggested using `srun --overlap` from
a second login shell:
```bash
srun --jobid=$JOBID --overlap nvidia-smi -l 2     # live snapshot
srun --jobid=$JOBID --overlap nvidia-smi dmon -s pucm -o T   # logged stream
srun --jobid=$JOBID --overlap nvidia-smi pmon -s um -d 2     # per-PID
```
Cross-reference fold PIDs to architecture via `grep "PID 12345"
output_kfold_v*.log`.

### 3c. First-run training diagnostic — upward prediction bias

Fold 9 epoch 11 output:
```
Min original_test_outputs: 14.85    Max: 185.08
Min original_test_targets:  0.84    Max: 120.00
Epoch 11  R²: -1.26  RMSE: 36.49
```

Predictions blew past the training cap (118 vs 120). Mean prediction was
~50 vs target mean ~22. Diagnosis: **KDE sampler at α=0.5 was creating a
distribution shift** — reweighted training distribution had heavier
upper-tail than the true distribution, pulling predictions up
systematically. Compounded by log target_transform's natural relative-
error weighting.

User experimented with α=0.2; results were "the same issue, drift seems
slower." Recommendation graduated to:

1. Drop KDE entirely (`--sampler-mode qcut --rebalance-min-ratio 0` =
   plain shuffle, no rebalancing).
2. Lower `--lr` from 2e-4 to 1e-4 for stability.

Rationale: **log target_transform alone handles SOC's log-normal scale**;
KDE rebalancing on top is partially redundant and introduces the
distribution-shift bias.

### 3d. "Violent" model-capacity reduction

Even without KDE, the model overfitted: training loss monotonically down,
test loss diverged from epoch 2 onward, predictions drifted up toward
the cap. Param count = 1.23M / 14.7k train rows = **84 params per
sample** — overfitting territory.

User asked: *"should I make a bigger 3DCNN? is it worth it?"*

Recommended NOT bigger 3DCNN — the architecture's three pool-2 layers
reduce 5×5×5 input to 1×1×1, destroying spatial information. Wrong
inductive bias for tiny patches.

Recommended instead: **reduce SGT capacity, hard**. Move to
`--model-size small` (SimpleSGT, 1 transformer layer, hardcoded) with
`--hidden_size 64` → 165 k params. Plus `--dropout_rate 0.5` for added
regularization. Plus number_of_epochs reduced (cosine decay over 60
instead of 300).

---

## 4. CRITICAL bug fix: `best_model_state` reference vs clone (`b5c1cac`)

**Discovery.** The user ran 8 spatial-CV sweeps under the new
small-model regime. Every config had negative mean R² across all folds.
Suspicion: maybe the gating + log target combo was fundamentally broken
at this scale.

**Diagnostic script** (`inspect_sweep.py`, committed `6a97d2f`) compared
the per-epoch R² trajectory in `fold_<i>_metrics.json` against the
"saved best" R² in `fold_<i>_summary.json`. Result was unmistakable:

| metric | every fold, every config |
|---|---|
| `saved R²` | exactly equal to `last epoch's R²` |
| `best epoch R²` (highest in the trajectory) | up to **+8 R² points higher** than `saved R²` |
| median gap | ~1.5 R² points |

**Root cause** in `SpatiotemporalGatedTransformer/train.py:337`:
```python
best_model_state = model.state_dict()
```

`model.state_dict()` returns a dict of references to the live parameter
tensors. The next epoch's optimizer step mutates those tensors in place,
so `best_model_state` always points at the most recent (final-epoch)
state. The "best" criterion was tracking the max R² correctly, but the
*weights* recorded as best were silently overwritten every subsequent
epoch.

**Fix:**
```python
best_model_state = {k: v.detach().clone()
                    for k, v in model.state_dict().items()}
```

**Impact.** This bug had been silently corrupting **every** kfold +
sweep + final-mapping output for the entire revision arc (and likely
the original submission as well). The corrected re-runs showed real
spatial-CV R² of 0.10–0.18 across architectures — a defensible result
under honest evaluation.

**Same fix ported to the three sibling architectures** in `b3b6814`:
- `3DCNN/train.py:223,263` (the `v.cpu()` pattern was accidentally
  correct on GPU but broken on CPU; made explicit)
- `CNNLSTM/train.py:237,241` (same root bug as SGT)
- `SimpleTransformer/train.py:252,256` (same root bug as SGT)

---

## 5. Sweep infrastructure (`6dd1812`, `029a8e5`, `41f7e34`, `eb90079`)

After the bug fix, the user wanted a systematic architecture sweep
rather than ad-hoc commands. Built a four-piece pipeline.

### 5a. `sweep_submit.py` — Slurm orchestrator

Submits one sbatch per architecture config (4-GPU jobs). Generates the
sbatch script on disk under `sweep/sbatch/<tag>.sbatch`, log files under
`sweep/slurm_logs/<tag>_%j.out`. Per-config output dir
`sweep/<sweep-name>/<tag>/`.

Initial grid (`6dd1812`): 8 SGT configs spanning d_model ∈ {48, 64, 96,
128}, heads ∈ {2, 4}, layers ∈ {1, 2, 3} with `--model-size big`. Reduced
to 80-epoch screening budget (later reduced again to 30 after diagnostic
work — see Phase 5d).

Added SimpleSGT rescue variants (`029a8e5`): `small_d64_h2_L1`,
`small_d96_h2_L1`, `small_d128_h2_L1` after early indication that the
big variants were over-parameterized.

Pivoted entirely to SimpleSGT (`41f7e34`): after the bug fix exposed that
SimpleSGT-class models hit R² ≈ 0.15–0.18, while the big EnhancedSGT
variants topped out around R² = −0.2 at 1.2M params. New default grid
became SimpleSGT-only, d_model ∈ {32, 48, 64, 96, 128}, h ∈ {2, 4}.

### 5b. `sweep_summarize.py` — ranking + comparison table

Recursive glob over `sweep/**/kfold_results_summary.json`, extracts
across-fold mean R² / std / RMSE / MAE / RPIQ + per-fold R² array.
Ranks by `score = r2_mean − 0.5 × r2_std` (rewards stability).

Tag regex handles `(small_)?d<H>_h<HEAD>_L<L>` for SGT,
`baseline_<model>_<variant>` for tree models, later extended to
`<family>_d<H>_h<HEAD>_L<L>` for the 3DCNN/CNNLSTM/SimpleTransformer
family ports.

### 5c. `--sweep-name` namespacing (`eb90079`)

Multi-axis sweeps (max-OC sensitivity, loss-function ablation) need to
land in separate directory trees so they don't overwrite. Added
`--sweep-name <name>` flag: outputs go to `sweep/<sweep-name>/<tag>/`
instead of `sweep/<tag>/`. `sweep_summarize.py` walks all sub-sweeps
recursively, surfaces a `group` column showing the namespace.

Used to run max-OC sensitivity:
- `--max-oc 90  --sweep-name oc90`
- `--max-oc 120 --sweep-name oc120`
- `--max-oc 150 --sweep-name oc150`

Finding: **R² is monotonically increasing in max-OC** (oc90: 0.10–0.14;
oc120: 0.14–0.17; oc150: 0.16–0.18) **for the same architecture**.
Driven by the heavy SOC tail's contribution to `ss_tot` in R²'s
variance-normalized denominator. Not a real model-quality signal — a
metric artifact. RPIQ moves the opposite direction (highest at oc90).
Drove a later careful discussion of RPIQ presentation (see Phase 11).

### 5d. Sweep diagnostic (`6a97d2f`)

After all 8 big-variant configs returned negative R², the user ran
the diagnostic script that exposed the best-state bug (Phase 4). The
diagnostic itself (`inspect_sweep.py`) walks the per-epoch
`fold_<i>_metrics.json` and reports `best epoch / saved R² delta`. After
the bug fix, delta is ≈ 0 everywhere as it should be.

---

## 6. Tree-ensemble baselines (`ef9930d`, `6afb8d3`)

**Goal.** Provide tree-ensemble baselines on the same 10-fold spatial-CV
geometry, same MAX_OC, same target transform — so the rebuttal table
shows SGT vs RF/XGB head-to-head under identical evaluation.

### 6a. `run_baselines.py` — GPU XGBoost + RF on the same folds

For each LUCAS sample, extract 80 features: `{mean, std, min, max}`
per band over the (T, H, W) cube, per the standard SOC-mapping
literature recipe (Hengl 2018, Padarian 2019). Tree models are
scale-invariant so no normalization. Targets log-transformed by default
to match the SGT runs.

XGBoost: native GPU support via `tree_method='hist', device='cuda'`.
Random Forest: cuML on GPU if installed; sklearn CPU otherwise (auto-
fallback). On 14.7k × 80, sklearn RF fits in ~30 s, so the GPU path is
nice-to-have not essential.

Folds reuse `build_folds_latitude_deciles` from `run_kfold.py`. Per-fold
outputs match SGT's format (`fold_<i>_predictions.parquet` +
`fold_<i>_summary.json` + `kfold_results_summary.json`) so they appear
in the unified `sweep_summarize.py` ranking.

### 6b. Sweep integration (`6afb8d3`)

Added `BASELINE_GRID` to `sweep_submit.py`: 4 XGB variants
(default/shallow/deep/fast — varying depth × n_estimators × lr) +
3 RF variants (default/shallow/deep — varying depth × n_estimators).
7 configs total.

Bundled into **one sbatch job** (1 GPU). Feature extraction cached to
`baseline_features/feats_max_oc_<C>.npz` so only the first config in the
bundle pays the ~3-min I/O cost; subsequent configs load the cache.

Avoided the race condition of submitting 7 parallel feature-extraction
jobs at once. New CLI: `--baselines` / `--no-baselines` /
`--baselines-only`, parallel to the later `--families` and `--vanilla`.

**Headline finding from baselines** (after bug fix): every RF and XGB
variant yields **negative R² on the Alpine fold (fold 0)** across all
max-OC values. SGT consistently positive there. This became one of the
three central claims of the revised paper (see Phase 10).

---

## 7. Cross-architecture sibling support — 3DCNN, CNNLSTM, SimpleTransformerV2 (`b3b6814`)

**Goal.** Make the 3DCNN, CNNLSTM, and SimpleTransformerV2 sibling
architectures runnable under the SGT k-fold pipeline so all four
families train on identical data, identical splits, identical augment-
ation, identical training loop.

### 7a. `--model-family` flag + generic model factory

Added `_build_model(args)` to `run_kfold.py` that dispatches:
- `sgt` → `build_sgt_model(args)` (the existing path)
- `3dcnn` → `Small3DCNN(input_channels, height, width, time, dropout)`
- `cnnlstm` → `RefittedCovLSTM(num_channels, lstm_input_size=128,
  lstm_hidden_size, num_layers, dropout)` (the 128 is hardcoded in
  the model's internal CNN — feature size after pool×2)
- `simpletransformer` → `SimpleTransformerV2(input_channels, height,
  width, time_steps, num_heads, num_layers, dropout_rate)`

All four take (B, C, H, W, T) input from the project-wide dataloader and
return a (B,) scalar. Sibling modules imported lazily inside the factory
so a default `--model-family sgt` run doesn't pay the import cost.

### 7b. `FAMILY_GRID` + `--families` / `--families-only` flags

Three sibling-family entries at d=64 h=4 L=1 (matched canonical config
across architectures). Submitted same way as SGT (one 4-GPU sbatch per
config). `sweep_summarize.py` regex extended to recognize
`<family>_d<H>_h<HEAD>_L<L>` tags.

### 7c. Findings

| family | best R² (oc150) | params | all folds R² > 0? |
|---|---|---|---|
| SGT (SimpleSGT d=128) | 0.168 | 363 k | **Yes** |
| SimpleTransformerV2 (d=64 h=4) | **0.183** | 11.2 M | Yes |
| CNNLSTM (d=64 h=4) | 0.090 | 93 k | No (fold 0 = −0.15) |
| 3DCNN | **−0.76** (catastrophic) | 27 k | No (every fold negative) |

3DCNN's failure: the architecture's three successive 2×2×2 poolings
reduce a 5×5×5 input to **1×1×1** — spatial information destroyed before
the FC head. Architectural mismatch, not capacity. Decided to leave
3DCNN as a "capacity-floor reference" in the table rather than fix it
(see Phase 11 "should we make a bigger 3DCNN" discussion).

---

## 8. Composite L1/MSE + Pearson chi-square loss (`a94941d`, `03c1e60`, `af1c6b4`)

**User request.** Add a chi-square loss component to SGT — for completeness
in the rebuttal, "for the sake of argument."

### 8a. First implementation (`a94941d`)

`_composite_loss()` helper in `train.py`. Five loss types:
- `l1`, `mse` — existing
- `chi2` — Pearson goodness-of-fit in original g/kg space, alone
- `l1_chi2` — L1 in transformed space + chi2 in original space
- `mse_chi2` — MSE in transformed space + chi2 in original space

Chi-square computed in **original space** (g/kg), not in the transformed
target space, because Pearson chi-square only has its statistical
interpretation on the natural scale of the target. For log-target runs
the helper inverts `exp(clamp(output, -3, 6))` before computing
`((y_pred - y_target) ** 2 / (y_target + eps)).mean()`. Clamp range
chosen so output ∈ [0.05, 403] g/kg — bounded even in early training.

Default `chi2_weight=0.01`.

### 8b. Redo with explicit α/β weights (`03c1e60`)

User's spec: `α × base + β × chi2`, with α and β as separate CLI knobs.
Renamed:
- `l1_chi2` → `composite_l1`
- `mse_chi2` → `composite_l2`

New flags: `--loss-alpha` (default 1.0) and `--chi2-weight` (default
0.1). Example: `--loss-alpha 0.5 --chi2-weight 0.1`.

Validated against the user-supplied spec by computing `0.5 × L1 + 0.1 ×
χ²` on synthetic data — matches arithmetically.

### 8c. Plumbed through the sweep submitter (`af1c6b4`)

Added `--loss-type` / `--loss-alpha` / `--chi2-weight` to `sweep_submit.py`,
forwarded to both `build_sbatch` (SGT) and `build_family_sbatch` (3DCNN,
CNNLSTM, SimpleTransformer). Default stays `l1` so existing sweep
configs are unchanged.

### 8d. Finding from composite_l2 runs

| variant | R² mean | std | notes |
|---|---|---|---|
| SGT `small_d128_h4_L1` oc150 plain L1 | 0.168 | 0.065 | rank 3 overall |
| SGT `small_d128_h4_L1` oc150 composite_l2 | **0.167** | **0.061** | rank 2 (best score) — marginal std improvement |
| SimpleTransformerV2 oc150 composite_l2 | **−0.110** | 0.251 | rank 82 — destabilized |

**SGT tolerates composite_l2 (marginal benefit on stability); the
canonical heavyweight transformer breaks under the same loss.** This
became an unexpected robustness data point for the SGT-vs-vanilla
comparison.

---

## 9. Vanilla transformer ablation (`1f3a3e1`, `045f246`)

**Motivation.** Pre-empt the strongest reviewer attack on the parameter-
efficiency claim: *"You compared SGT at 363k to SimpleTransformerV2 at
11.2M — that's a strawman. What about a parameter-matched vanilla
transformer?"*

### 9a. New model class (`1f3a3e1`)

`SpatiotemporalGatedTransformer/VanillaSpatiotemporalTransformer.py` —
**SimpleSGT minus the Gated Residual Network**. Replaces lines 23–36 of
`SimpleSGT.py` (`grn` + `gate` + `residual_proj` + `layernorm` of those)
with a single `Linear(feature_dim → d_model) + LayerNorm`. Everything
else (spatial encoder, positional embedding, transformer encoder,
output MLP head) is byte-identical to SimpleSGT.

Parameter counts at matched hyperparameters:

| Config | SimpleSGT | VanillaTransformer | Δ (vanilla − gated) |
|---|---|---|---|
| d=128, h=4 | 362,609 | **214,769** | **−147,840** (−41%) |
| d=96,  h=4 | 258,545 | 150,737 | −107,808 |
| d=64,  h=4 | 164,721 | 94,897 | −69,824 |

**The GRN block costs ~40% of SimpleSGT's parameter count at d=128.**

Registered as model family `vanilla_transformer` in
`run_kfold.py:_build_model`. Tag pattern is identical to other family
entries: `vanilla_transformer_d<H>_h<HEAD>_L<L>`.

### 9b. `--vanilla` flag separated from `--families` (`045f246`)

User wanted vanilla to be independently controllable. Pulled
`vanilla_transformer` entries out of `FAMILY_GRID` into a dedicated
`VANILLA_GRID` (two entries: d=64 and d=128, both h=4 L=1). New flags:
`--vanilla`, `--no-vanilla`, `--vanilla-only`, parallel to
`--baselines` / `--families`. Mutual-exclusion guards in `main()`
updated so each `*-only` flag truly excludes the other groups.

### 9c. Finding — gating doesn't earn its keep at this scale

| Config | Params | R² mean | R² std | Fold-0 R² |
|---|---|---|---|---|
| **`small_d128_h4_L1`** (gated SGT) | **363 k** | **0.170** | 0.062 | +0.09 |
| **`vanilla_transformer_d128_h4_L1`** | **215 k** | **0.170** | 0.069 | +0.11 |
| Δ (vanilla − gated) | −148 k | +0.0001 | +0.007 | +0.02 |

At d=64 the picture is **stronger for vanilla** (R² 0.168 vs SGT 0.156 —
vanilla wins on R² and is 70k params smaller).

**Verdict: the GRN does not provide measurable benefit at the LUCAS data
scale.** Vanilla is smaller, equivalently accurate, slightly better on
the Alpine fold. The single biggest re-framing of the paper falls out of
this finding (see Phase 10).

---

## 10. Final-models / production-mapping pipeline (`d160733`, `d5a0f1b`)

**Goal.** After the spatial-CV sweep established *evaluation* findings,
the rebuttal also needs **production maps**: each winning architecture
trained on the FULL dataset (no spatial holdout — that's evaluation
only) and applied to the Bavaria 1mil reference grid. Side-by-side
comparison figure becomes the rebuttal's headline visual.

### 10a. Five new scripts under `rebuttal/final_models/`

| Script | Role |
|---|---|
| `train_full.py` | Neural-network training on full data (95% train + 5% random monitor for best-state tracking — explicitly NOT spatial) |
| `train_full_baselines.py` | RF / XGBoost training on full data; reuses `extract_features_for_df` from `run_baselines.py` |
| `infer_bavaria.py` | Auto-dispatches between NN and tree paths based on checkpoint contents; runs over 1mil Bavaria grid for the requested year |
| `submit_finals.py` | sbatch generator: for NN families, 4-GPU train + dependent 1-GPU infer; for tree models, 1-GPU combined train+infer |
| `compare_maps.py` | Loads every map output, aligns on common grid, computes pairwise Pearson correlation + signed/abs differences, emits N-panel comparison figure |

### 10b. Path resolution fix (`d5a0f1b`)

Initial commit hardcoded `SOC_ROOT.parent / 'Weights-…/'` for Model A
checkpoint paths — worked on laptop, broke on cluster
(`/e/project1/scifi/fourel1/SGT/` has no `Weights-…` sibling).

Fixed: resolve via `SOC_WEIGHTS_DIR` from `_paths.py`. Added clear error
message at load time pointing at the env-var fix, the rsync option,
and the `--checkpoint` / `--analysis-pkl` CLI overrides — instead of
the bare `FileNotFoundError` stack trace.

### 10c. Expected Slurm cost

| Per-pipeline component | Time | Resources |
|---|---|---|
| NN train | ~1 h | 4 GPUs |
| NN infer (1M grid) | ~1.5 h | 1 GPU |
| Tree combined (train + infer) | ~2 h | 1 GPU |

With 12 pipelines (6 architectures × 2 band variants — see Phase 13) and
≥3 concurrent Slurm slots: **~3 h wall clock**.

---

## 11. Rebuttal-coverage scripts — the three gaps (`db8bd4b`)

After surveying `rebuttal/`, found that **23 of 26 Tier-2/3 action items
already had scripts** (`bootstrap_cis.py`, `split_comparison.py`,
`temporal_regression_corrected.py`, `extended_regression.py`,
`multi_run_cv.py`, `residual_sd_analysis.py`, `nn_distances.py`,
`figure7_replacement.py`). Wrote the three missing ones.

### 11a. `covariate_temporal_stats.py` — R2.M5

Annual statistics for the 5 dynamic covariates (LAI, LST, MODIS_NPP,
SoilEvaporation, TotalEvapotranspiration) at the 16,514 LUCAS sample
locations across 2007–2023. Linear trend test per band. Output:
`covariate_temporal_trends.png` + .md + .json.

**Argument for the rebuttal:** if covariates at sample locations are
temporally stable, the +0.751 g/kg/yr SOC coefficient cannot be a real
climate-driven trend the model detects in predictors — must be sampling
bias.

### 11b. `mc_dropout_uncertainty.py` — R3.9 + R4.4

Loads Model A, runs N=30 MC dropout forward passes on the validation set,
reports per-sample (mean, std, p05, p95, PI width, in-PI flag) and
aggregate 90% PI coverage. Two-panel spatial map (MC mean + MC std).

**Honest framing:** noted that observed coverage < 0.85 indicates MC
dropout under-dispersion (known property at moderate dropout rates per
Gal & Ghahramani 2016); deep ensembles cited as the principled
extension.

### 11c. `annual_soc_maps.py` — R2.M5

Model A inference over the 1mil grid at three target years: 2007 (start),
2015 (middle), 2023 (end). 2023−2007 difference map. Computes the
model-implied annual rate.

**Rebuttal value:** if model-implied annual rate is far below the
sample-distribution rate of +0.751 g/kg/yr, the temporal coefficient is
empirically a sampling-bias artifact rather than a real signal the
model detects.

---

## 12. Strategic rebuttal-framing discussions (no code)

Multiple back-and-forth discussions about how to present the substantively
changed paper.

### 12a. Is this OK as a revision (vs fresh submission)?

User: *"is it OK to return from revision with a totally different paper?
because this is what happens?"*

Verdict: **yes, with diplomatic framing.** The methodology improvements
were reviewer-requested. The findings (gating doesn't help, +0.751
trend is sampling bias) emerged from the rigorous experiments. The
dataset, study area, prediction target, and architecture family are
unchanged. The substantive change is the framing, driven by the
science.

Recommended:
1. Short heads-up email to Budiman Minasny ~1 week before deadline
2. Cover letter that **leads with the diplomatic framing** ("we
   conducted reviewer-requested experiments; they led us to refine the
   contribution; here's what changed")
3. Three things explicitly retracted:
   - Chi-square loss component (became a no-op due to internal magnitude
     rescaling — honest acknowledgment is better than retraining)
   - +0.751 g/kg/yr temporal coefficient (reframed as
     sampling-distribution diagnostic)
   - The G in SGT (gated residual network — ablation shows it doesn't
     help at this scale; recommend the gateless variant)
4. Title shift options ranked by aggressiveness; favored **option 2:
   "Lightweight Transformers for High-Resolution Soil Organic Carbon
   Mapping: A Comparative Study Under Spatial Cross-Validation."**

### 12b. Comparison with Padarian 2019 (canonical DSM-CNN paper)

User pasted Padarian et al. 2019 ("Using deep learning for digital soil
mapping," SOIL journal) and asked how the SGT results compare.

Key context: Padarian reports R² ≈ 0.51–0.64 on topsoil under 90/10
random splits with bootstrap OOB. SGT reports R² ≈ 0.17 under 10-fold
spatial CV with 1.2 km buffer.

Direct comparison is misleading because of **evaluation protocol**.
Cited Roberts et al. 2017 (*Ecography*) + Ploton et al. 2020 (*Nature
Communications*): spatial CV systematically compresses R² by 2–3×
relative to random splits, with adequate buffer. Applying a 2.5×
compression factor to Padarian's 0.51 gives a predicted spatial-CV R²
of ~0.20 — **exactly where SGT lands.**

The two papers are at the same operational benchmark, just under
different evaluation rigour. SGT additionally provides: honest spatial
CV protocol, comparison against canonical baselines, parameter-
efficiency (122k vs Padarian's ~unknown but likely 100k+ for the 2-conv
multi-task CNN), 4× more training data.

### 12c. "Lightweight transformer" reframing

User: *"basically this can be a study on how a lightweight transformer
is the most versatile architecture on this type of problem"*

**Confirmed.** Data supports this exactly:

| Dimension | Lightweight transformers (gated or vanilla) | Heavy transformer | Trees | Tiny CNN |
|---|---|---|---|---|
| R² mean (oc150) | 0.16–0.18 | 0.18 (≈ tied) | 0.12–0.14 | −0.76 |
| All folds R² > 0 | **Yes** (across configs) | Yes | **No** (fail Alpine) | No |
| Fold-0 R² (Alpine extrapolation) | +0.04 to +0.12 | +0.07 | −0.07 to −0.41 | −0.74 |
| Parameters | 100k–400k | 11.2 M | n/a | 27 k |
| Robust to max-OC cap (90/120/150) | Yes | Yes | Worse at oc90 | No |
| Robust to architectural sub-variants | Yes (d=32→128, h=2/4, gated/vanilla) | n/a (single config) | n/a | n/a |
| Robust to loss function (L1/MSE/composite_l2) | Gated: yes; vanilla: mostly | **No** (breaks on composite) | n/a | n/a |
| Band-extensible (O(1) in C) | Yes (per-band 1×1 input) | **No** (O(C²)) | n/a (handcrafted features) | Yes |

**Six axes of robustness for lightweight transformers; every other
family fails at least one.** That's "most versatile" empirically.

### 12d. RPIQ caveat

User asked how to explain low RPIQ (0.66–0.88 vs conventional
"poor < 1.4" cutoff). Three caveats to surface in the rebuttal:

1. RPIQ benchmarks (Bellon-Maurel et al. 2010) come from spectroscopic
   SOC studies under random splits on controlled lab samples — not
   directly comparable to spatial CV.
2. RPIQ is cap-sensitive: same architecture goes from RPIQ=0.88 at
   max-OC=90 to RPIQ=0.70 at max-OC=150 because the heavy tail's
   contribution to RMSE outpaces its contribution to IQR.
3. Within a fixed protocol and cap, **RPIQ ordering matches R²
   ordering**: SGT > SimpleTransformerV2 ≈ tied; both > all trees.

Recommended: footnote-style disambiguation, not a full paragraph; the
absolute value reflects honest evaluation, not deficiency.

### 12e. "Should we make a bigger 3DCNN?"

Verdict: **not worth it.** The 3DCNN failure is architectural (the
three pool-2 layers collapse 5×5×5 → 1×1×1), not capacity. Bigger
channels won't fix the spatial-information loss. Leaving 3DCNN as
"capacity-floor demonstration" in the table is actually useful — shows
that "lightweight" alone isn't enough; you also need the right
inductive bias. Fixing it could even hurt the story if a tuned 3DCNN
matches SGT at smaller scale (would dilute the architectural-
contribution claim).

Pre-emptive paragraph drafted for the rebuttal: *"Our naive 3D-CNN
baseline (Small3DCNN, 27 k parameters, three pool-2 layers) was not
competitive (R² = −0.76) due to architectural mismatch: three
successive 2×2×2 poolings reduce the 5×5×5 patch to a 1×1×1
representation, eliminating spatial information before the FC head. We
retained this configuration as a capacity-floor reference; a redesigned
3D-CNN preserving spatial structure would be expected to reach the R²
range of our tree-ensemble baselines (~0.10), still below SGT."*

### 12f. SOTA framing

Five claims are publishable as-is:
1. SGT outperforms standard tree-ensemble baselines under spatial CV.
2. SGT is the only architecture with positive R² on every spatial fold.
3. SGT is parameter-efficient (92× smaller than SimpleTransformerV2 at
   matched R²).
4. SGT is band-extensible (O(1) in C vs O(C²) for SimpleTransformerV2).
5. SGT under the SAME pipeline as every baseline — methodologically
   clean comparison.

"SOTA" without qualifiers is overreach (would invite "but did you
compare against Padarian / Hengl / Wadoux?" reviewer attacks). Defensible
scoped version:

> "SGT is, to our knowledge, the only architecture tested under strict
> spatial CV that combines transformer-class accuracy with sub-200k-
> parameter efficiency and demonstrated extrapolation robustness on the
> Alpine regime."

---

## 13. 6-band vs 20-band ablation (`21f44c4`)

**Last substantive change.** User flagged that the 6 → 20 band expansion
was *not* reviewer-requested and was substantive (different model
inputs, different parameter scaling, different absolute R²). Needed to
disentangle the band-expansion contribution from the architectural
contribution.

### 13a. `band_subsets.py` — single source of truth

`get_band_indices(name, full_list)` returns indices into `bands_list_order`.
`band_suffix(name)` returns `'_6band'` or `'_20band'` for auto-tagging.

Original 6 = `{Elevation, LAI, LST, MODIS_NPP, SoilEvaporation,
TotalEvapotranspiration}` — the original-submission covariate set, sits
at indices 0–5 of `bands_list_order` by construction (Phase 3a).

### 13b. `--bands-list` flag plumbed everywhere

Added to:
- `run_kfold.py` — `_BandSubsetWrapper` slices channel dim AFTER
  `_NormalizingWrapper` (so the project-wide 20-channel feature_means/
  stds stay valid). `_build_model` uses `len(band_indices)` for
  `input_channels`.
- `run_baselines.py` — 80-d feature vector is sliced to the relevant 24
  columns when `--bands-list=original_6` (4 stats per band × 6 bands).
  Cache file is shared across runs at the same max-OC; the slicing
  happens after cache load.
- `sweep_submit.py` — `--bands-list` forwarded to run_kfold and
  run_baselines; sweep_name auto-appends `'_6band'` so 6-band and
  20-band sweeps don't collide.
- `train_full.py` + `train_full_baselines.py` — flag forwarded; run_name
  auto-suffixed with `_6band` / `_20band` so checkpoints don't collide.
- `infer_bavaria.py` — reads bands_list from saved checkpoint config and
  replays the same slicing at inference time (both NN and tree paths).
- `submit_finals.py` — `--bands-lists` CSV flag default
  `'full_20,original_6'` runs **both** variants of every config. Doubles
  the work (8 NN pipelines + 4 tree pipelines per band variant = 12
  total).
- `compare_maps.py` — `DEFAULT_RUNS` auto-expanded across both band
  suffixes.

### 13c. Rationale documented in the cover letter

Three converging justifications for the band expansion:
1. **scorpan-model coverage** (McBratney et al. 2003): original 6 covered
   c, o, r; revised 20 adds **s** (Clay, Sand, pH, BulkDensity, CEC) +
   densifies c (Precipitation, AirTemperature, SoilMoisture, SnowDepth)
   and r (Slope, Aspect, TWI). Soil-property dimension was *missing*
   from the original submission.
2. **Empirical band-extensibility demonstration**: 6→20 band ratio
   confirms architecture's O(1) parameter scaling (SGT +1.2%) vs
   SimpleTransformerV2's O(C²) scaling (+548%). Not a hypothesis
   anymore — an empirical result on the actual data.
3. **Alignment with operational DSM standards**: SoilGrids 2.0 uses
   400+ covariates; Padarian 2019 used 5 (acknowledged limitation);
   20 is modest by current standards.

Citation Minasny will appreciate: he is the M in McBratney/Mendonça/
Minasny 2003. Original scorpan author.

### 13d. Cost

NN training compute doubles: 8 NN jobs × 2 band variants = 16 (was 8
before this phase). Tree pipelines: 4 tree jobs × 2 = 8 (was 4). Total
24 sbatch jobs for the full final-models comparison matrix. ~6–7 h wall
clock at 4 concurrent slots.

---

## Summary — final state of the pipeline

### Spatial-CV evaluation pipeline (`rebuttal/gpu_experiments/spatial_kfold/`)

| Component | Purpose |
|---|---|
| `run_folds_parallel.py` | Slot-based orchestrator, `--folds-per-gpu` packing |
| `run_kfold.py` | Per-fold training; KDE/qcut sampler; D4 augmentation; composite loss; band-subset support |
| `run_baselines.py` | RF/XGBoost on per-band statistics; same fold geometry |
| `sweep_submit.py` | Sbatch generator with three independent grids: SGT (`DEFAULT_GRID`), siblings (`FAMILY_GRID`), vanilla ablation (`VANILLA_GRID`), tree baselines (`BASELINE_GRID`) |
| `sweep_summarize.py` | Recursive ranking across sweep namespaces |
| `inspect_sweep.py` | Per-epoch trajectory diagnostic (originated the best-state bug discovery) |
| `band_subsets.py` | `--bands-list` resolution helper |

### Production-mapping pipeline (`rebuttal/final_models/`)

| Component | Purpose |
|---|---|
| `train_full.py` | NN training on full data, no spatial holdout |
| `train_full_baselines.py` | RF/XGBoost on full data |
| `infer_bavaria.py` | 1mil-grid inference, auto-dispatches NN/tree |
| `submit_finals.py` | Sbatch generator; `--bands-lists` CSV defaults to both variants |
| `compare_maps.py` | Cross-architecture comparison figure (auto-expanded across band variants) |
| `README.md` | Workflow doc |

### Rebuttal analyses (`rebuttal/`)

| Script | Reviewer addressed |
|---|---|
| `bootstrap_cis.py` | R1.3, R3.6 (bootstrap CIs) |
| `split_comparison.py` | R3.3 (random vs spatial split distributional stats) |
| `temporal_regression_corrected.py` | R1.2, R3.5 (temporal coefficient sensitivity) |
| `extended_regression.py` | R3.5 (land-use multivariate regression) |
| `multi_run_cv.py` | R1.3, R3.6 (repeated CV evidence) |
| `residual_sd_analysis.py` | R3-mod11 (training residual SD anomaly) |
| `nn_distances.py` | R2.M3 (nearest-neighbour distance distribution) |
| `figure7_replacement.py` | R2-m2 (2D replacement for the 3D scatterplot) |
| `covariate_temporal_stats.py` | R2.M5 part 1 (covariate temporal stability) |
| `mc_dropout_uncertainty.py` | R3.9, R4.4 (uncertainty quantification) |
| `annual_soc_maps.py` | R2.M5 part 2 (multi-year inference maps) |

### Architecture comparison table — empirical findings (oc150)

| Model | Params | R² mean | R² std | Fold-0 R² | All folds R² > 0? |
|---|---|---|---|---|---|
| **VanillaTransformer (proposed, d=128)** | **215 k** | **0.170 ± 0.07** | 0.069 | **+0.11** | **Yes** |
| SimpleSGT (gated, d=128, ablation reference) | 363 k | 0.170 ± 0.06 | 0.062 | +0.09 | Yes |
| SimpleSGT (d=48, h=2) | 122 k | 0.178 ± 0.09 | 0.089 | +0.04 | Yes |
| SimpleTransformerV2 (canonical heavyweight) | 11.2 M | 0.183 ± 0.09 | 0.087 | +0.07 | Yes |
| CNNLSTM (d=64, h=4) | 93 k | 0.090 ± 0.10 | 0.100 | −0.15 | **No** |
| 3DCNN (Small3DCNN) | 27 k | −0.76 ± 0.35 | 0.354 | −0.74 | **No** |
| Random Forest (best) | n/a | 0.135 ± 0.10 | 0.100 | −0.10 | **No** |
| XGBoost (best) | n/a | 0.140 ± 0.15 | 0.155 | −0.26 | **No** |

---

## Commit timeline

| SHA | Phase | Title |
|---|---|---|
| `feb2a8d` | 1 | parallel orchestrator: --folds-per-gpu |
| `fb58def` | 2 | rebuttal kfold: defensible cap, KDE sampler, augmentation, stratified report |
| `25ae836` | 3a | SGT config: extend bands_list_order 6 → 20 |
| `6dd1812` | 5a | sweep: architecture grid via Slurm |
| `029a8e5` | 5a | sweep: add SimpleSGT rescue variants |
| `6a97d2f` | 5d | inspect_sweep diagnostic |
| `b5c1cac` | **4** | **train.py: clone tensors when capturing best_model_state (critical fix)** |
| `41f7e34` | 5a | sweep: pivot grid to SimpleSGT neighborhood |
| `ef9930d` | 6a | baselines: GPU XGBoost + RF on same 10-fold splits |
| `6afb8d3` | 6b | sweep_submit: bundle RF + XGBoost baselines |
| `eb90079` | 5c | sweep: --sweep-name namespacing |
| `a94941d` | 8a | composite L1/MSE + Pearson chi-square loss |
| `03c1e60` | 8b | composite loss redo: explicit α/β weights |
| `b3b6814` | 7 | 3DCNN / CNNLSTM / SimpleTransformer under SGT's pipeline |
| `af1c6b4` | 8c | sweep_submit: plumb composite loss flags |
| `db8bd4b` | 11 | covariate temporal stats + MC dropout UQ + annual SOC maps |
| `d5a0f1b` | 10b | rebuttal scripts: SOC_WEIGHTS_DIR path fix |
| `d160733` | 10a | final_models: production-mapping pipeline |
| `1f3a3e1` | 9a | vanilla_transformer: SimpleSGT minus the GRN |
| `045f246` | 9b | sweep_submit: split vanilla into own grid + flag |
| `21f44c4` | 13 | band_subsets: --bands-list flag everywhere |

---

## Outstanding / next steps

### Code

1. **Smoke-test the 6-band SGT pipeline on cluster** before launching the
   full 24-job matrix. Single command:
   ```bash
   python rebuttal/final_models/train_full.py \
       --run-name smoketest_sgt_6band --bands-list original_6 \
       --model-family sgt --model-size small --hidden_size 64 \
       --num_heads 2 --num_layers 1 --dropout_rate 0.5 \
       --lr 1e-4 --lr-scheduler cosine --lr-min 1e-6 \
       --loss_type l1 --target_transform log --max-oc 150 \
       --per-gpu-batch-size 256 --effective-batch-size 256 \
       --num-epochs 2 --seed 42 --augment-train
   ```
   Should print `Model: SimpleSGT family=sgt (162,705 trainable params)`.

2. **Run the full 24-job matrix** once smoke-test passes:
   ```bash
   python rebuttal/final_models/submit_finals.py
   ```

3. **After everything lands**, generate the rebuttal-ready comparison
   figure:
   ```bash
   python rebuttal/final_models/compare_maps.py
   ```

### Writing

1. **Cover letter** with the diplomatic framing (Phase 12a structure):
   - Opening paragraph acknowledging scope-of-revision and offering
     editorial flexibility
   - Three explicit retractions (chi-square loss, +0.751 trend, gating)
   - Per-reviewer point-by-point response
2. **Heads-up email to Budiman Minasny** ~1 week before deadline
   (around 2026-05-22).
3. **Title shift**: option 2 ("Lightweight Transformers for
   High-Resolution SOC Mapping…").
4. **Tier-1 editorial items** still pending from the action plan: T1.8
   (Figure 14 → Table), T1.9–T1.14, T1.17, T1.18, T1.22–T1.26.

### Strategic

- **Don't run more architecture experiments.** The 24-job final-mapping
  matrix + the spatial-CV sweep is enough to support every claim.
- **Drop the chi-square loss component from the abstract**; keep as
  supplementary ablation.
- **Lead with the parameter-efficiency + extrapolation-robustness
  combination.** That's the rebuttal-winning pair.

---

*Log maintained on `bestrun-bands`. Last updated 2026-05-19.*