Local decision models with calibrated probabilities. Send a state (text or JSON) and typed questions (yes/no, choice, score); get a probability for every option back. No text generation, no network calls, no per-call cost. The API follows the Jev-style Decisions format, so existing clients mostly need only a new base URL.

Model	gutsy-0.8b v0.4, a fine-tune of Qwen3.5-0.8B: Q8_0 (775 MB) or Q4_K_M (505 MB, 25-30% faster), on Hugging Face
Runtime	gutsy-inference, llama.cpp on an ordinary CPU
Docs	Model card, release notes
License	Apache 2.0

Why use it

• Calibrated. A 0.8 means roughly 80% (validation calibration error 0.022).
• Local and private. Runs on CPU; your data never leaves the machine.
• Deterministic. Same request, same probabilities, on the same machine and build.
• Robust to option order. 96-97% of answers are unchanged when options are shuffled.
• Cheap follow-ups. The state is encoded once; each extra question costs only its own tokens.
• Up to 255 options, and a reject probability for "none of these fits".

Quickstart

git clone https://github.com/kouhxp/gutsy
pip install -e gutsy/gutsy-inference           # needs llama-cpp-python >= 0.3.35
pip install -U huggingface_hub
hf download kouhxp/gutsy --include "gutsy-0.8b-v04*" --local-dir gutsy/gutsy-inference/models
cd gutsy/gutsy-inference
cp models.example.json models.json
gutsy-inference check                   # expect "caching OK" for both models
gutsy-inference serve                   # http://127.0.0.1:8765

curl -s http://127.0.0.1:8765/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": "Order 1182 shipped Monday. The customer says it arrived Thursday with the box crushed.",
  "questions": {"decision": {"type": "choice", "instructions": "What should support do next?",
    "criteria": {"refund": "issue a refund", "replace": "send a replacement",
                 "escalate": "a manager reviews it within 24 hours"}}}}'

The default model is Q8_0; send "model": "gutsy-0.8b-v04-q4" for the smaller file. Choice and score answers also return confidence, margin and reject. Endpoints, question types and how to set thresholds are in the runtime README.

Browser demos

Start the server with gutsy-inference serve --cors and open any file in gutsy-inference/examples. The data is made up; every probability comes from your local model.

Support triage	Feedback sheet	Model router
Team, refund and mood for one ticket	Ask any yes/no question of 30 rows	Pick the cheapest model per message
Fact check	Call radar	Maze autoplay
Score each claim against a source	Live triggers on a sales call	gutsy plays a maze chase

--cors lets any page you visit query the server, so leave it off when you're not using the demos.

Results

All gutsy numbers are ours, on public items; scores are accuracy. Other systems' numbers come from the sources noted and may use different samples, so treat comparisons as indicative.

gutsy vs Laya vs Jev

	gutsy Q8_0	Laya	Jev
JevBench public items (231)	0.732	0.537	0.866
BANKING77, zero-shot over all 77 intents	0.613	0.382	0.764

JevBench: fine-tuned Laya and Jev as reported by other projects on the same public set; official JevBench results use sealed items. BANKING77: gutsy on the full test split; Laya and Jev from dhruvmehra/jevbench on a 500-item sample of it.

JevBench by tier

	Q8_0	Q4_K_M
Easy (48)	48	47
Original (72)	65	62
Hard (111)	56	52
Total	0.732	0.697

Benchmarks

	gutsy Q8_0	gutsy Q4_K_M	Jev
Financial PhraseBank	0.927	0.926	0.770
RAGTruth *	0.813	0.807	0.773
WinoGrande *	0.620	0.629	0.907
JudgeBench	0.539	0.563	0.786
BBH	0.481	0.474	0.943

* Training splits of these were in the training data (test items never were). Jev figures are from the Jeff-0.8B project's README, on a different sample of these benchmarks.

Measured by us on the same 700 items per task:

	gutsy Q8_0	Jev (via API)
BoolQ	0.837	0.906
SNLI	0.870	0.883
CommonsenseQA (not trained on)	0.577	0.863

Also: lists of 20-254 options 0.991; abstains when a missing fact decides the answer 0.896.

Where it is weak

Test on your own data before automating any of these:

• Date and number arithmetic is near chance. Compute in code and ask about the result.
• New multi-step rulebooks (JevBench long_policy 0.32).
• Ambiguity: still hedges in about a third of cases where the rules already settle it.
• Judging close pairs of answers and world knowledge (see JudgeBench and CommonsenseQA).
• Calibration on hard, unfamiliar items (error 0.137, vs 0.022 on validation).
• Long states on CPU: median 4.3 s, 95th percentile 27 s on JevBench's hard tier (Q8_0).
• English only.

Training

Full fine-tune of Qwen3.5-0.8B on about 105,000 questions (one epoch, one 80 GB GPU, about an hour), with cross-entropy plus Brier loss on soft labels. A third of the data is scenarios written by an open teacher model (Qwen3.8-27B) and kept only when two independent passes agreed; the rest is public datasets (reading, classification, routing, judging, contracts) and code-generated puzzles with exact labels. The build fails if any question template's answer can be predicted from surface features such as option length or position. Details and sources: MODEL_CARD.md.

Disclosures

• No closed-model output is used anywhere in the training data.
• Some puzzles target known weak spots on BBH and JevBench families, but no BBH or JevBench item, template or wording was used, and training data was decontaminated against JevBench and the evaluation panel.
• In-domain data: RAGTruth, WinoGrande, CLINC150 and MASSIVE training splits. BANKING77 is held out, so the comparison above is zero-shot for all three systems.
• No official JevBench result yet.

License

Code and weights: Apache 2.0, as are the base and teacher models. The training data uses CC BY 4.0, CC BY-SA, CC0, MIT and Apache 2.0 sources, credited in MODEL_CARD.md (Jev Decisions v1 is CC BY 4.0 with upstream NVIDIA terms); redistributing the data itself means following each source's terms.