# Cortex Agent: kodiak

> Open encoder-only decision model from Cortex Agent: a state plus typed choice and score questions in, calibrated answers out in one forward pass, with a per-question "can't tell" probability. 1.04B parameters on Ettin-encoder-1B; research-stage, English only.

- Page: https://systemonemodels.ai/cortex-agent-llc/kodiak
- API: https://api.systemonemodels.ai/v1/models/cortex-agent-llc/kodiak
- Download: `pip install systemonemodels && systemone pull cortex-agent-llc/kodiak`

## Facts

| | |
|---|---|
| Maker | Cortex Agent (https://systemonemodels.ai/cortex-agent-llc) |
| Decides | choice, score, classify, route |
| Architecture | kodiak |
| Base model | jhu-clsp/ettin-encoder-1b |
| Parameters | 1.0B |
| Licence | apache-2.0 |
| Availability | Open weights |
| Released | 2026-10-06 |
| Latest version | 0.3.0 |

## Reported evaluation

Suite: Kodiak frozen eval set v0.2, never-seen choice tasks, forced answer (Cortex Agent's own eval; this seed-1 checkpoint). Numbers are the publisher's own.

- Decision accuracy: 67.5%
- Calibration error (ECE): 0.074

## Model card

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

# Kodiak-v0.3-1B

Open encoder-only decision model from Cortex Agent: a state plus typed choice and score questions in, calibrated answers out in one forward pass, with a per-question "can't tell" probability. 1.04B parameters on Ettin-encoder-1B; research-stage, English only.

Created by Shane Larson. Ettin-encoder-1B (MIT) with choice, score and abstain heads; the state can be text, a list of texts or JSON, and choice labels are set per request. There is no typed yes/no question; "can't tell" is abstention. v0.3 adds seven decision kinds taught with synthetic data (pairwise judge, sarcasm, written-policy violation, long-answer hallucination, stance, refund eligibility, agent-step safety) but, on the maker's own eval, does not improve never-seen accuracy over v0.2. Limits the maker states: reworded options give the same answer only about two thirds of the time on never-seen tasks, implied policy violations are often missed, arithmetic is near chance, ratings are rough, and requests over about 8,000 tokens are refused. The maker's Qwen3-8B comparison (0.688 never-seen, calibration error 0.293) is its own baseline run. Non-commercial and share-alike training sources were excluded. Siblings include a three-model accuracy mode, v0.2 and smaller ModernBERT previews. A hosted API is only on a waitlist. The third-party Decision Index lists v0.2 accuracy mode near the bottom of its board.

## What it decides

- **choice** — picks one option from a set
- **score** — places the input on an ordered scale
- **classify** — assigns a category from a fixed taxonomy
- **route** — sends the input to one of several destinations

## At a glance

| | |
|---|---|
| Parameters | 1.04B |
| Base model | `jhu-clsp/ettin-encoder-1b` |
| Maker | Cortex Agent |
| Released | 2026-10-06 |
| License | apache-2.0 |
| Reported accuracy | 67.5% |
| Reported latency | about 38 ms per request on a GPU; a few hundred ms on a CPU (maker's figures) |

## Get the weights

```bash
pip install systemonemodels
systemone pull cortex-agent-llc/kodiak
```

The files are served from the maker's Hugging Face repository, [`cortex-agent-llc/kodiak-v0.3-1b`](https://huggingface.co/cortex-agent-llc/kodiak-v0.3-1b), and verified against the checksums recorded here.

## Read more

- [Model card](https://huggingface.co/cortex-agent-llc/kodiak-v0.3-1b)
- [Accuracy mode](https://huggingface.co/cortex-agent-llc/kodiak-v0.3-1b-accuracy)
- [kodiak-small-v2-preview](https://huggingface.co/cortex-agent-llc/kodiak-small-v2-preview)
- [Code and fine-tuning kit](https://github.com/grizzlypeaksoftware/kodiak)

---

*This page was opened by System One for Cortex Agent, 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