amihai4by/logic-reasoner-v2

Logic Reasoner v2

Logic Reasoner v2 is a verdict-style reasoning and verification model packaged for Ollama and distributed as GGUF. It is designed for operational, infrastructure, and automation workflows that require deterministic, machine-readable output, not conversational text.

---

Why this model exists

Most large language models are optimized for human conversation, not for systems that must act on model output.

In operational environments this causes recurring issues:

• Non-deterministic phrasing that breaks parsers
• Excess verbosity that hides the actual decision
• Missing information that is not explicitly surfaced
• Explanations instead of decisions

Logic Reasoner v2 exists to address this gap.

It enforces a strict reasoning interface on top of a general-purpose language model by:

• Requiring a clear verdict: true, false, or uncertain
• Separating reason, evidence, assumptions, and next actions
• Guaranteeing exactly one JSON object
• Explicitly stopping generation after the structured response

This makes the model suitable as a decision and verification component inside automated systems, not just as a chat assistant.

---

What this model is for

Use this model when you need:

• A clear verdict instead of a narrative
• Structured reasoning that can be logged or audited
• Predictable output suitable for automation
• A bridge between LLM reasoning and operational workflows

Typical use cases:

• Kubernetes and GPU stack troubleshooting (GPU Operator, DCGM, drivers)
• Verification of technical or operational claims
• Incident triage and post-mortem workflows
• JSON-driven automation pipelines

---

What this model is not for

This model is not intended for:

• Academic benchmark leaderboards (e.g. MATH500, GSM)
• Strict symbolic math grading
• Creative or open-ended generation
• Long conversational interactions

---

Output contract

When used with the provided Modelfile, the model outputs exactly one JSON object and then stops.

Schema

{ "verdict": "true | false | uncertain", "reason": "string", "confidence": 0.0, "evidence": ["string"], "assumptions": ["string"], "next_actions": ["string"] }

Rules

confidence is a heuristic value between 0.0 and 1.0

If information is missing, the verdict must be uncertain

No text outside JSON is expected when the wrapper is used

Stop behavior is enforced by the Modelfile

How to run with Ollama

Create the model locally:

ollama create logic-reasoner-v2 -f Modelfile

Example request:

curl http://localhost:11434/api/generate -d '{ "model": "logic-reasoner-v2", "stream": false, "prompt": "Input: DCGM exporter reports 0 GPUs across all nodes. Question: Is the system healthy?" }'

Quantization

Format: GGUF

Quantization: Q4_K_M

Optimized for low-latency operational inference

Provenance

This model was built and packaged as part of the LLM FUN project on NVIDIA DGX B200 infrastructure using:

Kubernetes (RKE2)

Ollama

OpenWebUI

The Modelfile is a core part of the model behavior and must be used to reproduce the intended output guarantees.

Limitations Confidence values are heuristic, not statistically calibrated

The base model may default to explanatory text if the wrapper is not used

Determinism applies to structure, not factual correctness

License MIT