Module 10 Activity

Scenario

Vision-language models answer everything, including questions the image cannot support.

What you build

A set of checkable prompts, a measured assessment of the three weak points, and computable refusal conditions.

Steps

1. Rewrite five open prompts as checkable ones, each with an explicit way to say the model cannot tell.
2. List six properties your images cannot show and ask about each. Record whether the model refused or answered.
3. Measure counting, small-text reading and spatial relations on 50 of your own images. Report the rate for each.
4. Require a region with every visual claim and validate it against the image bounds.
5. Implement the refusal conditions that are computable before the model call - region size and blur.
6. Separate the policy refusals into routing, not the prompt.

Evidence to hand in

• Five rewritten prompts with explicit unsure options.
• The six unobservable questions and the model's behaviour.
• Measured rates for counting, small text and spatial relations.
• The region requirement with bounds validation.
• The computable refusals and the routed policy refusals.

Review checklist

• Every prompt has an explicit way to say the model cannot tell.
• The three weak points are measured on your own images, not assumed.
• Regions are validated against image bounds.
• Computable refusals happen before the model call.
• Policy refusals are in routing, not in the prompt.

Mark complete