Academy
2. The Five Controls Explained
Identify the five independent checks—source anchoring, prompt fidelity, action claims, code guard, and repetition—and understand what each one verifies.
Every AI answer on OUPI is independently evaluated by TrustLayer through five controls. Each control checks a different dimension of reliability, and they work independently—one can pass while another raises a warning. Together they produce a single score out of 100 and a color (green, orange, or red). Understanding what each control actually verifies helps you interpret the badge and know exactly where to look when something is flagged.
1. Source Anchoring — When the AI answer draws on documents, knowledge bases, or web results, this control checks whether the claims in the answer are actually found in those sources. If the AI states a fact that doesn't appear in any provided source, source anchoring flags it. Click the badge to see exactly which claims were matched and which were not. Note: if no sources are involved, this control simply has nothing to anchor against.
2. Prompt Fidelity — This control verifies that the answer covers what you asked and respects the constraints you set. Did you ask for three bullet points and get a paragraph? Did you request a summary of section 2 but receive content from section 5? Prompt fidelity catches these mismatches. It ensures the AI stayed on task relative to your actual request.
3. Action Claims — Sometimes the AI says it performed an action: "I sent the email," "The file has been saved," "I searched the database." This control compares those claimed actions against the tool calls that actually happened. If the AI says it did something but no corresponding action was executed, this control raises a red flag. This is one of the most critical checks—an uncaught false action claim could have real consequences.
4. Code Guard — When the AI generates code, this control scans it for dangerous patterns. It acts as a safety net before you copy-paste or execute anything. If a code snippet contains risky constructs, code guard will warn or flag the answer so you can review the code carefully before using it.
5. Repetition — This control detects when the model is recycling the same content rather than producing a genuinely useful response. Repeated or looping text is a known failure mode of language models, and this guard ensures you are alerted when the output is not adding real value. Like the other controls, it contributes independently to the overall score.
Remember: a green badge means the answer is consistent with the sources, your request, and the tools used—it does not guarantee the answer is true in the real world. Your judgment still matters, especially for decisions and commitments. Also, short conversational replies (like "sure, which one?") are marked as not applicable rather than scored, because there is nothing to verify.
TrustLayer runs five independent checks on every AI answer: (1) Source anchoring — are claims found in the provided sources? (2) Prompt fidelity — does the answer match your request and constraints? (3) Action claims — did claimed actions actually happen? (4) Code guard — is generated code safe? (5) Repetition — is the model recycling content? Each contributes to the score out of 100. Click the badge to see per-control details, and always apply your own judgment on top.