# Provenance

This repo is a curated export of the `rrm` research workspace. `trm/`, `hrm/`, `srm/` are
upstream clones with our modifications committed as plain files (nested `.git` not included).

| dir | upstream | pinned commit | our changes |
|---|---|---|---|
| `trm/` | github.com/SamsungSAILMontreal/TinyRecursiveModels | `c01103738605ba39d1430519b1ee0c62f4c707f8` | `pretrain.py` +625 lines: trajectory-augmentation training (multi-rollout perturbation, directional FD variants). Backward-compatible — `trajectory_augment` defaults False, so the official Maze/Sudoku recipes run unchanged. |
| `hrm/` | github.com/sapientinc/HRM | `ac15626f8db096a63c775b84c9dc868776a6feda` | `pretrain.py` +383 lines: same trajectory-augmentation additions. |
| `srm/` | github.com/YurenHao0426/srm (fork of sapientinc/HRM) | tracked separately at that remote | Stable Recursive Model (our architecture). |

To diff against upstream: clone the upstream repo at the pinned commit and diff `pretrain.py`.

Excluded from this export (regenerable / heavy): all `checkpoints/`, `wandb/`, `outputs/`,
`__pycache__/`, `*.pt`, `*.npz`, `*.png`, the Sudoku dataset (rebuild via
`scripts/build_datasets.sh sudoku`). The Maze dataset IS included for clone-and-run.