qwen3.5-9b-structured-review-profile

Qwen3.5 9B — Structured Review Profile

Configuration and evaluation artifact — not a trained or fine-tuned model. This repository contains no model weights. It documents a bounded way I evaluated and used Qwen/Qwen3.5-9B for advisory review workloads in the Winwood AI Toolkit.

Why this exists

Small local models become more useful when the surrounding system owns the decision boundaries. This profile packages the part around the model:

• exact-output JSON contracts for three review tasks;
• explicit allowed and unsafe decisions;
• prompt rules that treat supplied evidence as the only authority;
• deterministic validation before any result reaches a person; and
• a human-controlled, canary-only promotion boundary.

The model proposes. Code validates. A person decides.

Workloads

Machine-readable settings and output contracts are in structured-review-profile.json, with sanitized prompt contracts in PROMPTS.md. The public companion dataset is mattwinwood/production-ai-guardrail-evals.

Historical local evaluation

These measurements were produced on 2026-08-29 through an OpenAI-compatible local gateway against a digest-pinned Ollama build named qwen3.5:9b. They describe that pinned local runtime, prompts, policies, and private source corpus together—not the upstream model in every environment.

The catalog workload did not meet the 90% promotion threshold. Every workload remained canary-only. The public dataset is a sanitized derivative and is not claimed to reproduce these numbers.

Intended use

Use this repository as a reference for designing structured, advisory model workflows where deterministic code and human authority remain outside the model. It is suitable for evaluation and system-design study, not drop-in production inference.

Limitations

• No weights, adapters, tokenizer, or inference endpoint are included.
• No fine-tuning was performed.
• The prompts and schemas do not make model output factual or safe by themselves.
• Results may change with a different quantization, runtime, prompt, policy, sampler, or model revision.
• This profile must not be used to authorize deployments, mutate user data, or make consequential decisions autonomously.

License

The original configuration and documentation in this repository are Apache-2.0 licensed. The upstream Qwen model remains governed by its own repository and license.