---
license: cc-by-nc-4.0
language:
  - en
library_name: gguf
pipeline_tag: text-classification
base_model: prism-ml/Ternary-Bonsai-2-27B-gguf
tags:
  - jev
  - jev-style
  - openjev
  - decision-model
  - typed-decisions
  - gguf
  - bonsai
  - ternary
  - local-inference
  - blackwell
---

# Jev-Bonsai-Compass

A complete, local **Jev-style text decision** download: [Prism's Ternary Bonsai 2 27B](https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf), a decision adapter derived from [OpenJev](https://huggingface.co/openjev/openjev), the tested native server, and a small Python decision API. This is an independent project, not an official Jev or OpenJev release. Fast choices use one short option-scoring pass without reasoning; numerical questions can use a checked CPU calculation.

## What was measured

| Local configuration | Correct on the same 720 text questions | Sampled GPU memory | Mean text-decision latency |
|---|---:|---:|---:|
| **Jev-Bonsai-Compass**, adapter + compact runtime | **611/720 (84.9%)** | **7,844 MiB**, one RTX 5080 | **177 ms** |
| Full OpenJev FP8 reference | 615/720 (85.4%) | 30,974 MiB total, two GPUs | 118 ms |
| Untuned Bonsai + the same compact runtime | 584/720 (81.1%) | 7,510 MiB, one RTX 5080 | 145 ms |

The 720 questions comprise 120 each from AG News, Emotion, BoolQ, SNLI, ARC-Challenge and a **20-intent subset** of Banking77. They used identical prompts and option order across the three local configurations. Current versus full OpenJev differs by four answers; a paired bootstrap interval for the accuracy difference is −2.50 to +1.25 percentage points. This does **not** establish formal parity, and it is **not** OpenJev's published 10,000-question or hosted Jev evaluation. The full reference used a different two-GPU serving stack, so its latency is not a same-hardware speed comparison. See [domain and method evidence](evidence/core-public-domains.json) and [benchmark metadata](benchmark-manifest.json).

Accuracy varies by task: this adapter scored 117/120 on science, 113/120 on the selected banking intents, **102/120 on SNLI**, and **65/120 on Emotion**. It scored 327/328 on reused, authored numerical regressions with compact calculator assistance; that is a system result, not raw-model arithmetic accuracy. [Numeric cases](benchmarks/numeric-regression.jsonl) and [summary evidence](evidence/core-current.json) are included. All timing and memory figures are serial local measurements with context 4,096 and one active request.

## Download and run

The complete repository includes one canonical Bonsai GGUF, the adapter, Python source and wheel, a Linux x86-64 native server and shared libraries, examples, licenses, and a [SHA-256 manifest](MANIFEST.json). The bundled native build targets NVIDIA Blackwell `sm_120a` and needs a driver plus CUDA 13, cuBLAS 13 and NCCL 2 system libraries. Python 3.10+ is required; the decision core itself has no third-party Python dependencies.

```bash
hf download ajh-code/Jev-Bonsai-Compass --local-dir Jev-Bonsai-Compass
cd Jev-Bonsai-Compass
python3 validate_release.py
bash run.sh serve --gpu 0 --host 127.0.0.1 --port 9304
```

In another terminal:

```bash
curl -s http://127.0.0.1:9304/v1/decide \
  -H 'Content-Type: application/json' \
  -d '{"state":"There are 9 boxes with 12 pens in each.","question":"How many pens are there?","options":[["a","108"],["b","96"],["c","21"]]}'
```

The result contains the selected option ID, method, timing and trace. `bash run.sh decide --input examples/requests.jsonl --gpu 0` processes JSONL without exposing HTTP. `bash run.sh serve --profile base --gpu 0` selects the untuned, Apache-licensed Bonsai path. The server has no authentication; use `--host 127.0.0.1` or a trusted network. For other GPU architectures, build a compatible [Prism llama.cpp fork](https://github.com/PrismML-Eng/llama.cpp) and pass `--binary /path/to/llama-server`.

## Scope and terms

The API accepts text `state`, `question`, and 2–26 `[id, description]` options at `POST /v1/decide`; it also provides `/health` and `/v1/capabilities`. It is **not** a drop-in `/v1/systemone` implementation, vision model, general chat model, or autonomous agent service. Option-relative scores are not guarantees of correctness. The separate experimental agent evidence is included for transparency but its controller is not part of this decision API.

The bundled adapter contains locally modified OpenJev-derived weight factors and retains **CC BY-NC 4.0** noncommercial terms. The unmodified Bonsai base weights are Apache 2.0, this Python software is Apache 2.0, and the native runtime is MIT. The [NOTICE](NOTICE), [licenses](licenses/) and [profile identities](profiles.json) distinguish the components. The adapter is GGUF-formatted data stored as `adapters/broad-lr0.0001.adapter`; `run.sh` loads it directly. The single `.gguf` file is the canonical base model used by both profiles.