README.md

---
license: apache-2.0
language:
- en
task_categories:
- text-classification
pretty_name: System One Mini Synthetic Diagnostics
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.jsonl
  - split: validation
    path: data/validation.jsonl
  - split: calibration
    path: data/calibration.jsonl
  - split: development_renderer
    path: data/development_renderer.jsonl
  - split: test
    path: data/test.jsonl
---

# System One Mini Synthetic Diagnostics

Deterministic synthetic controlled-intervention summaries used by `DavidHatley/system-one-mini`. The records contain no real coding-agent traces, private repositories, personal data, external model outputs, or teacher labels.

**This dataset and model are independent research artifacts, not reproductions of Jev or RLCD.**

## Splits

| Split | Rows |
|---|---:|
| `train` | 20,000 |
| `validation` | 2,000 |
| `calibration` | 2,000 |
| `development_renderer` | 2,000 |
| `test` | 2,000 |

`development_renderer` is the former sealed renderer split that was subsequently used to compare candidates. It must not be described as an unbiased final test. `test` is a separate renderer written and frozen after model selection, then evaluated once for the initial release.

## Fields

- `id`: SHA-256 fingerprint of the latent synthetic situation
- `state`: rendered diagnostic summary
- `labels`: five numeric labels in schema order
- Five named string label fields
- `latent`: complete generated situation used to compute labels
- `provenance`: generator, split, renderer, and seed

## Generation

Generation code and manifests are included in `generator/` and `provenance/`. The policy makes labels mechanically identifiable: every component trial starts independently from the same original failing snapshot, and exactly one clearing intervention identifies the operational failure source.

## Limitations

- All scenarios are synthetic and templated.
- The label taxonomy covers only four component classes plus `unknown`.
- Recurring logical patterns across splits do not represent natural incident prevalence.
- Public test labels make the split unsuitable as a permanently sealed benchmark for future model selection.