---
license: mit
language:
  - en
task_categories:
  - other
tags:
  - agents
  - world-models
  - planning
  - tool-use
  - synthetic
  - reproducible-research
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files: mini_tasks.jsonl
---

# MiniOS — a deterministic symbolic task suite for tool-agent world models

This dataset is the **MiniOS** task suite from **nanoAWM**. Code, the world
model trained on it, and the full paper live in the GitHub repository:
**https://github.com/jlov7/nanoAWM**.

MiniOS is a deterministic, fully symbolic toy operating system: filesystem,
terminal/tests, git, browser/forms, DB/table, email/calendar, package/config,
approvals, hidden state, irreversible actions, delayed consequences, and
rollback. Each task is constructed to exhibit **consequence aliasing** — two
candidate actions that look equally reasonable from the current visible
observation but diverge sharply in future reward, reversibility, hidden-state
corruption, or approval compliance. A memoryless reactive policy provably cannot
be optimal across the hidden states; a history-conditioned consequence model
can.

## Contents

`mini_tasks.jsonl` — **420 tasks across 15 families**, deterministically
generated (seeded).

| Split | Tasks | Purpose |
| --- | ---: | --- |
| `train` | 240 | training (used with 3 vocabulary-randomized variants each → 720) |
| `val` | 30 | validation / calibration only |
| `test` | 60 | in-distribution evaluation |
| `template_holdout` | 45 | in-distribution evaluation (see disclosure below) |
| `ood` | 45 | local out-of-distribution stress |

The 15 families include approval-gated email send, irreversible file deletion,
delayed test failure, database-migration hidden invariant, git dirty-tree,
rollback recovery, package-install dependency conflict, cross-surface
dependency, ambiguous same-visible-observation, and adversarial failure
variants.

### Schema (one JSON object per line)

| Field | Meaning |
| --- | --- |
| `task_id`, `template_id`, `seed` | identity / generation provenance |
| `family`, `split` | task family and split membership |
| `description` | the natural-language task given to the agent |
| `visible_state` | the observation the agent sees (surface, object, phase, …) |
| `actions` | the candidate action set |
| `hidden_state` | environment ground truth (approval status, corruption, …) |
| `reactive_action` | the action a memoryless reactive policy would pick |

**`hidden_state` is environment ground truth, not a planning-time input.** It is
needed to *run* the environment and score outcomes, but the learned planner's
feature extractor explicitly excludes it (along with task/split ids, labels,
rewards, and oracle next-state). Shipping it here lets you reproduce the
environment; it is not a leak into the agent.

## Honest disclosure: what these splits are, and are not

This dataset is **partly lexically separable** and the eval splits **share
vocabulary with training**. Treat the split names literally and skeptically:

- Measured feature-level overlap with a training task: **`test` 90%,
  `template_holdout` 100%, `ood` 71%**. The split named `template_holdout` does
  **not** hold out templates.
- Treat "held-out 1.000" in the paper as *in-distribution* performance. The
  honest generalization measurement rebuilds the suite over a **disjoint
  vocabulary** (held-out object names *and* disjoint nonce action markers
  sharing no tokens with training) — on which the learned planner scores
  **0.524** vs a 0.067 best baseline and a 1.000 oracle.
- The safe/risky distinction is partly encoded in action strings, so a pure
  keyword matcher with no world model already scores ~0.73 here. A large part of
  the task is lexical, not consequence reasoning.

This is local symbolic evidence for studying consequence modeling under hidden
state — **not** a web/OS benchmark and **not** external validation.

## Usage

```python
import json
tasks = [json.loads(line) for line in open("mini_tasks.jsonl")]
train = [t for t in tasks if t["split"] == "train"]
```

Regenerate deterministically (CPU-only, no network) from the GitHub repo:

```bash
python -m nanoawm.envs.generate --suite mini --out data/mini_tasks.jsonl
```

The disjoint-vocabulary generalization suite (the source of the honest 0.524
number) is produced by `python -m nanoawm generalization_ladder --run-ladder`.

## Citation

```
@software{nanoawm,
  title  = {nanoAWM: a tiny action-conditioned world model for tool agents},
  author = {Lovell, Jason},
  year   = {2026},
  url    = {https://github.com/jlov7/nanoAWM}
}
```

License: MIT.

## Personal project disclaimer

<sub>Personal, independent research and development. Not affiliated with, endorsed by, or representative of any employer, client, or organization.</sub>