# Sthānika AI: sieve

> Sthānika AI's open decision-model family. A LoRA adapter and answer head on a Qwen3.5 backbone read a text or JSON state once and return a calibrated probability for every option of choice, yes/no and score questions, without generating text.

- Page: https://systemonemodels.ai/sthanika-ai/sieve
- API: https://api.systemonemodels.ai/v1/models/sthanika-ai/sieve
- Download: `pip install systemonemodels && systemone pull sthanika-ai/sieve`

## Facts

| | |
|---|---|
| Maker | Sthānika AI (https://systemonemodels.ai/sthanika-ai) |
| Decides | choice, score, noul |
| Architecture | sieve |
| Base model | qwen/qwen3.5-9b |
| Parameters | 9.0B |
| Licence | apache-2.0 |
| Availability | Open weights |
| Released | 2026-09-29 |
| Latest version | 2026.09 |

## Reported evaluation

Suite: LocalLLaMA/typed-decisions test split (2,000 decisions; Sthānika AI's own run of Sieve-9B). Numbers are the publisher's own.

- Decision accuracy: 70.4%
- Calibration error (ECE): 0.040

## Model card

<!-- generated by scripts/seed_catalog.py; edit content/models/catalog.yaml -->

# Sieve-9B

Sthānika AI's open decision-model family. A LoRA adapter and answer head on a Qwen3.5 backbone read a text or JSON state once and return a calibrated probability for every option of choice, yes/no and score questions, without generating text.

This entry is Sieve-9B: a rank-32 LoRA (two rank-16 runs averaged) and a Kev-style pointer head on the post-trained Qwen3.5-9B, trained mostly on Kev's decision-v7 data. The repo holds the adapter and head, not the base weights; it loads through the sieve package from GitHub and needs about 20 GB of GPU memory. Each question runs from the cached state; up to 255 options. On the maker's development partitions: 87.2% in-domain (ECE 0.028), 82.9% on an out-of-domain transfer set. Sthānika's own run of the third-party Decision Index 0.2.1 kit gives 41.71; the board's 0.3 release shows Sieve 9B at 41.88 and Sieve 2B at 22.16. Siblings, built differently: Sieve-4B (a 255-way option-code head on Qwen3.5-4B; 67.7% on typed-decisions, self-run Index 43.82, not on the board) and sieve-2b (scores each option separately on 16 blocks of Qwen3.5-2B-Base, so option order cannot change the answer). Weights are Apache-2.0, but part of the training data keeps its own terms, some non-commercial (27% of Sieve-4B's records); the maker asks users to check them before commercial use. The cards say the lab is not affiliated with TypeSafe AI.

## What it decides

- **choice** — picks one option from a set
- **score** — places the input on an ordered scale
- **noul** — answers a yes/no question with one probability

## At a glance

| | |
|---|---|
| Parameters | 9B |
| Base model | `qwen/qwen3.5-9b` |
| Maker | Sthānika AI |
| Released | 2026-09-29 |
| License | apache-2.0 |
| Reported accuracy | 70.4% |

## Get the weights

```bash
pip install systemonemodels
systemone pull sthanika-ai/sieve
```

The files are served from the maker's Hugging Face repository, [`sthanika-ai/Sieve-9B`](https://huggingface.co/sthanika-ai/Sieve-9B), and verified against the checksums recorded here.

## Read more

- [Model card](https://huggingface.co/sthanika-ai/Sieve-9B)
- [Sieve-4B](https://huggingface.co/sthanika-ai/Sieve-4B)
- [sieve-2b](https://huggingface.co/sthanika-ai/sieve-2b)
- [Code and data](https://github.com/sthanika-ai/Sieve)

---

*This page was opened by System One for Sthānika AI, who can claim the organisation and take it over at any time.*

---

From System One Models — https://systemonemodels.ai/ · every System One model: https://systemonemodels.ai/system-one-models