# T2 Residue Classifier

This is a Tier 2 experiment using `output_base: "p"`. The model emits a single
base-p digit, so the task is a learned residue classification problem rather
than fixed-width decimal digit generation.

Inference-time preprocessing reduces each operand separately modulo `p`, matching
the representation normalization used by the reference neural baselines. The
model then uses learned p/residue embeddings, an MLP scorer, and a learned
low-rank bilinear residue-product head to emit logits over residues `0..255`.
It does not compute `(a*b) mod p` at inference time in Python or tensor code.

Train locally:

```powershell
.\.venv\Scripts\python.exe .\my-t2-model\train.py --minutes 10
```

GPU full-table continuation:

```powershell
.\.venv\Scripts\python.exe .\my-t2-model\train.py --minutes 8 --resume --full-table --batch 8192 --bilinear-dim 128
```

Evaluate locally:

```powershell
.\.venv\Scripts\modchallenge.exe check .\my-t2-model
.\.venv\Scripts\modchallenge.exe evaluate .\my-t2-model --total 110
.\.venv\Scripts\modchallenge.exe evaluate .\my-t2-model --total 1100
```

For a minimal HuggingFace submission, keep `manifest.json`, `model.py`,
`weights.pt`, and this README. `train.py` is development-only.