decision-1

decision-1 is a decision adapter for Qwen3.8-27B: a LoRA adapter plus a pointer head. It reads a state and typed questions and returns a probability for every option, by prefill alone, without generating text. It is packaged as a .ninfer adapter for NInfer-4090, which serves it on two wire protocols from the same process and weights as chat:

• OpenAI Decisions, POST /v1/decisions, for the OpenAI Python SDK 3.26.0 or newer.
• TypeSafe System One, POST /typesafe/v1/systemone, for the TypeSafe SDK 0.6 and 0.7.

It is not a PEFT, Transformers, or GGUF file.

File

Field	Value
Filename	decision-1.lora.ninfer
Size	89,716,300 bytes
SHA-256	14015d187d244a37dcc1fc0593308a756b8be1c1b70914bf088188acfb341a39
Kind	decision adapter
LoRA	rank 16, alpha 32, on query, output gate, key, value, attention output, Gated DeltaNet output, and MLP down projection
Pointer head	256-dimensional query and key projections over the final-norm hidden state of width 5,120
Temperature	1.414

Use it

The adapter needs NInfer-4090 and the Qwen3.8-27B NInfer artifact. Download it into the directory that the server reads with --lora-dir. Decisions are qualified on BF16 KV:

hf download tensorninja2/decision-1 decision-1.lora.ninfer --local-dir lora
./build-sm89/apps/ninfer-serve models/qwen3_8_27b.ninfer --kv-dtype bf16 \
  --max-context 32768 --lora-dir lora --lora-slots 2

OpenAI Decisions:

from openai import OpenAI

client = OpenAI(api_key="local", base_url="http://127.0.0.1:8080/v1")
result = client.decisions.create(
    model="decision-1",
    input="I was charged twice this month. Please refund one charge before Friday.",
    questions=[{
        "type": "predicate", "name": "billing",
        "instructions": "Is this message about billing?",
    }],
)
print(result.answers[0].probability)  # 0.939

TypeSafe System One, where jev-latest resolves to the server's only decision adapter:

from typesafe_sdk import Noul, TypeSafeClient

client = TypeSafeClient(api_key="local", base_url="http://127.0.0.1:8080/typesafe")
result = client.system_one(
    state="I was charged twice this month. Please refund one charge before Friday.",
    questions={"billing": Noul(instructions="Is this message about billing?")},
)

On an RTX 4090 with a cached state, one question takes about 25 ms and six questions about 100 ms. See Decisions for both protocols, latency, and limits.

Best with Caudra

Caudra, the coding agent from the author of NInfer-4090, can use decision-1 as its local decision engine, in the same process and weights as the chat model. Caudra uses it for permission advice, Auto mode screening, shell effect predictions, content screening of tool results, subagent routing, and reranking, and its decisions.decide() and decisions.rerank() calls work inside code snippets. Add the endpoint to ~/.config/caudra/providers.toml:

[decisions.ninfer]
protocol = "openai"
base_url = "http://127.0.0.1:8080/v1"
model = "decision-1"

Then select it with /decision-model or caudra decisions use ninfer. The full setup, including the chat provider, is in Use it from Caudra.

Evaluation

Scored with BF16 KV on the clean knowable questions of the decision-v7 corpus from kev. The locked test partition was scored once.

Split	Questions	Accuracy	Brier	NLL	ECE
Development	1,264	0.871	0.184	0.356	0.024
Locked test	1,200	0.856	0.190	0.355	0.020

Kev-27B, which adapts every linear layer and trained on a superset of this data, publishes 0.866 and 0.870 accuracy on the same splits. On the test split, score questions are the weakest type at 0.596 accuracy, against 0.913 for choice and 0.929 for noul.

Training

QLoRA and the pointer head on an NF4-quantized Qwen3.8-27B base, one epoch on the decision-v7 training split, with temperature 1.414. Training ran in the llm-datasets repository. tools/convert/qwen3_8_27b/convert_lora.py --decision-head converted the PEFT directory and head to .ninfer.

Limits

• Decisions meet kev's per-question tolerance on bf16 and int8 KV. The rotated KV modes, including the 262K-context rk4v4, miss the probability bar on a few questions.
• The adapter was trained on text. NInfer accepts images in OpenAI Decisions, but no image-task accuracy or calibration is established for this adapter.
• There is no refusal detection, and the adapter does not reproduce the outputs or calibration of OpenAI's or TypeSafe's hosted decision models.

License

Apache-2.0 for the adapter and head. The Qwen3.8-27B base is Apache-2.0. The datasets behind the decision-v7 corpus carry their own licenses; see kev's model cards.