Academy
4. Why Some Answers Are Not Scored
Understand why short conversational replies are marked as not applicable and why this keeps the badge meaningful.
TrustLayer evaluates every AI answer against five independent controls — source anchoring, prompt fidelity, action claims, code guard, and repetition — and produces a score out of 100 plus a color badge. But not every reply warrants that analysis. When you type "Hey, can you help me?" and the AI responds "Sure, what do you need?", there are no factual claims, no cited sources, no actions performed, and no code generated. Running five controls against an empty set of verifiable content would produce a meaningless number. That is why TrustLayer marks these short conversational replies as "not applicable" instead of assigning a score.
The "not applicable" label is a design choice that protects the badge's credibility. If every casual exchange received a green 100/100, users would learn to ignore the badge — it would look like it always passes. By withholding a score when there is nothing to verify, TrustLayer ensures that when you do see a green, orange, or red badge, it reflects a genuine evaluation. Green means the answer's claims align with visible sources and your request. Orange means at least one control flagged something worth checking. Red signals a serious issue. Each color carries weight precisely because trivial replies are excluded.
What counts as a "short conversational reply"? Think of messages that contain no claim about the world, no reference to a document, no assertion that an action was taken, and no generated code. Examples: greetings, clarifying questions ("Which file do you mean?"), or simple acknowledgments. The moment the AI makes a verifiable statement — quotes a figure from your knowledge base, says it sent an email, or produces a code snippet — the controls kick in and a score appears. The boundary is not about word count; it is about whether there is something concrete to check.
If you see "not applicable" on a reply that clearly contains claims or data, click the badge to inspect the control details. This can reveal whether the AI's response was treated as conversational when it shouldn't have been, and you can rephrase your prompt to get a scored, verifiable answer.
Remember: even a scored green badge does not guarantee truth. TrustLayer compares the answer against what it can see — your sources, your prompt, the tools called. A statement consistent with those inputs but factually wrong in the real world can still pass. The badge lowers risk; it does not replace your judgment. By keeping unscored replies out of the mix, TrustLayer makes sure the scores you do receive are as informative as possible.
Open a chat in OUPI and send a simple greeting like "Hello, how are you?" Look at the AI's reply: you should see the badge marked as not applicable. Then ask a factual question grounded in one of your documents and compare — this time a score and color should appear. Click the badge to explore the control breakdown.
Short conversational replies — greetings, clarifications, acknowledgments — contain no verifiable claims, so TrustLayer marks them as not applicable rather than inflating scores. This keeps the badge meaningful: when a score does appear, it reflects a real evaluation across five controls. Green, orange, and red carry weight because trivial exchanges are excluded. And even with a score, the badge is a reading aid — not a certificate of truth. Your judgment still matters.