Keeping an AI feature honest

Second placeholder post. Same rule as the first: this structure is yours, the words are not — replace them with a real story from your own work.

The promise is the product

Start with the claim your feature makes to the user. Then write down what happens when the model gets it wrong. If that second answer does not exist yet, the feature is not ready.

Sources and limits

Write which documents, tables or APIs the model may read, and what it must never invent. State the boundary in language a product owner can check, not in model jargon.

The fallback

Describe what the interface does when confidence is low: show the source, ask the user to confirm, or refuse cleanly. A clear refusal is a feature; a confident wrong answer is a bug.

Checklist before you ship

• The user can tell when an answer came from a model.
• Every generated claim can be traced to a source you control.
• There is a manual path that does not use the model at all.

Replace this scaffolding with your own words before the site goes public.