the model that doesn't write: an intro to decision models

Description and use

Decision models return numeric probabilities for questions, enabling programmatic branching, unlike chatbots that output text.

Problem

Need a way to get numeric decision outputs (probabilities) from a model for use in code branches, rather than textual chatbot responses.

Outcome

The workflow demonstrates that decision models provide actionable numeric outputs, enabling automated decision logic and reducing reliance on human review for clear cases.

Workflow

• 1. Define a JSON schema with questions (type: noul, choice, score).
• 2. Send image and schema to the decision model via a single API call.
• 3. Receive structured probabilities and confidence for each question.
• 4. Apply branching logic (e.g., if dog_present < 0.5 reject, if breed confidence > 0.9 auto_tag, else send to human).
• 5. Render decisions visually with a custom Python function in a Gradio workflow.
• 6. Deploy the workflow as a single API endpoint using Gradio Client.

Evidence

you get back a probability for every allowed answer

Source

if answers["dog_present"]["noul"] < 0.5: reject(photo) # 0.98 — skipped

Source