# Anonymous KAFT MVP reproduction

This repository is a minimal, self-contained reproduction of two core
observations for deep message-passing GNNs:

1. a plain deep GCN can lose usable backward transport even when a
   standardized hidden-state probe remains informative; and
2. Kronecker-Aligned Feedback Training (KAFT) improves training while leaving
   the forward GCN unchanged.

The package intentionally contains only the code needed for this MVP. It does
not include manuscript files, review material, cached datasets, private paths,
or repository history.

## One-click notebook

Open `reproduce_mvp.ipynb` and choose **Run All**. The first cell installs the
local package, the experiment downloads Cora automatically, and the remaining
cells train BP and KAFT, run the gradient diagnostic, display the result
tables, and save machine-readable artifacts.

The same notebook can be executed non-interactively:

```bash
python -m pip install -r requirements.txt
jupyter nbconvert \
  --to notebook \
  --execute reproduce_mvp.ipynb \
  --inplace \
  --ExecutePreprocessor.timeout=1200
```

## Command-line reproduction

```bash
python -m pip install -e .
python run_mvp.py
```

The default CPU protocol uses:

- Cora with the public Planetoid split;
- an identical six-layer, width-64, bias-free GCN forward model for BP and
  KAFT;
- seeds 0, 1, and 2;
- Adam, learning rate 0.01, weight decay \(5\times10^{-4}\), 200 epochs;
- KAFT diffusion \(\alpha=0.5\), 10 fixed linear propagation steps;
- hop cap \(K=3\), 64 Gaussian probes, and alignment every 10 steps;
- a separate ten-layer, 100-epoch BP diagnostic.

KAFT changes only the hidden-layer backward rule. Its graph-side feedback is
\(P_\ell(\hat A)D(\hat A)\), and its feature-side matrix is aligned to the
chain-normalized suffix-weight probe target. The output layer uses the true
cross-entropy error.

## Outputs

Running either entry point writes:

```text
artifacts/
├── mvp_results.json
├── mvp_summary.csv
└── training_curves.png
```

The JSON file contains the full configuration, every seed-level result,
training histories, and the ten-layer gradient diagnostic. Test accuracy is
reported at the validation peak; the test labels are never used for
selection.

## Expected behavior

Exact numbers can vary slightly with library versions and hardware. The
checked-in executed notebook and artifacts record the environment used for
the release. The intended qualitative checks are:

- KAFT exceeds BP test accuracy in the six-layer comparison;
- the ten-layer BP diagnostic reaches exact-zero weight gradients;
- the output-adjacent preactivation error remains finite; and
- the standardized penultimate hidden-state probe remains above chance.

The checked end-to-end CPU run produced:

| Method | Test accuracy at validation peak |
|---|---:|
| BP | \(68.87\pm1.20\%\) |
| KAFT | **\(78.53\pm1.33\%\)** |

For the separate ten-layer BP diagnostic, all ten weight gradients were
exactly zero in 3/3 seeds. The output-adjacent error remained between
\(1.26\times10^{-4}\) and \(1.60\times10^{-4}\), while the standardized
penultimate probe obtained 51.4--57.8% accuracy (seven classes).

## Scope

This is a compact verification path, not the complete experimental suite.
It covers the diagnostic and the central BP-versus-KAFT mechanism on one
standard benchmark. The full evaluation uses additional datasets, backbones,
depths, normalizers, and ablations.