---
license: cc-by-4.0
pretty_name: Budgeted Guardrails for MLP Channel Interventions
tags:
  - preprint
  - reproducibility
  - transformers
  - language-modeling
  - pytorch
  - apple-mps
  - training-dynamics
---

# MLP Channel Guardrails

This repository accompanies the Zenodo preprint
[`10.5281/zenodo.21003379`](https://doi.org/10.5281/zenodo.21003379),
**Budgeted Guardrails for MLP Channel Interventions in Small Transformer
Language Models**.

The study evaluates training-time MLP-channel interventions in small causal
Transformer language models. The supported claim is narrow: in the tested
small-model settings, low-amplitude MLP channel gates with strict event budgets
reduce trajectory damage relative to fixed event gating and dropout while
keeping the trained model dense at inference time.

The project no longer claims that targeted dormant-channel rejuvenation works.
Matched-random controls are as strong or stronger than dormant-targeted
interventions in the current evidence.

## Release Artifacts

This Hugging Face dataset repository is a public mirror of the Zenodo preprint
release package.

Primary release files:

- `output/pdf/budgeted_guardrails_mlp_channel_interventions.pdf`
- `output/zenodo/budgeted-guardrails-mlp-channel-interventions-20260617.zip`
- `paper/budgeted_guardrails_mlp_channel_interventions.tex`
- `paper/references.bib`
- `paper/reproducibility_manifest.md`
- `paper/zenodo_metadata.json`

Recommended citation metadata and artifact details are in
[`README_ZENODO.md`](README_ZENODO.md).

## Repository Contents

The release package includes:

- Small GPT-style causal Transformer training code.
- MLP-channel telemetry and channel-gating intervention code.
- Launch, aggregation, statistical testing, and figure-generation scripts.
- Generated diagnostics and manuscript tables under `reports/` and
  `paper/tables/`.
- Public bounded text-data caches and metadata under `data/raw/`.
- Dependency-free tests under `tests/`.

Paper reproducibility details are in
[`paper/reproducibility_manifest.md`](paper/reproducibility_manifest.md).

Training experiments must use MPS only. If MPS is unavailable, stop before
model creation. CPU, CUDA, and `PYTORCH_ENABLE_MPS_FALLBACK=1` runs are invalid
for this project.

## Verification

Dependency-free tests:

```bash
PYTHONPATH=src python -m unittest discover -s tests -v
```

MPS gate:

```bash
PYTHONPATH=src .venv/bin/python scripts/check_mps.py
```

Paper consistency check for result-table, claim-language, bibliography,
figure-reference, CI-artifact, abstract-length, and section-order drift:

```bash
PYTHONPATH=src .venv/bin/python scripts/check_paper_consistency.py
```

First smoke run, only after the MPS gate passes:

```bash
PYTHONPATH=src .venv/bin/python -m dormant_rejuvenation.train \
  --run-name smoke_telemetry \
  --out-dir runs/exploratory \
  --steps 50 \
  --eval-every 25 \
  --telemetry-every 5
```

Bounded corpus sample download:

```bash
python scripts/download_text_sample.py \
  --name tinystories_sample \
  --url https://huggingface.co/datasets/roneneldan/TinyStories/resolve/main/TinyStories-train.txt \
  --out data/raw/tinystories_sample.txt \
  --max-bytes 8000000
```