Stage 7: ground an AI and measure the answers

Outcome: an evaluation scorecard. Time: 1 to 2 hours.

Explain and inspect

Grounding gives an AI selected evidence and asks it to stay inside it. Retrieval chooses context; generation writes an answer. Good retrieval cannot repair a false source or guarantee faithful output. Start with a clear prompt and relevant bundle link. Model Context Protocol (MCP) earns its extra infrastructure only when bounded search, identity and logging improve measured results. Read the grounding evidence.

Start with a worked question

Use the fictional context prompt and expected answers or the government evidence exercise. Choose a supported access route and first identify exactly what your AI received. Explorer's local MCP adapter remains a retrieval prototype, so include an MCP comparison only when you have a verified compatible server.

Do and check

Freeze the candidate and questions. Write the expected identity, required claims, sources and serious errors. Run the same held-back questions with no context, direct prompt-and-link context and bounded MCP retrieval if available. Require citations and allow “not supported”. Do not send confidential sources to an unapproved service.

Use the scorecard. Count identity selection, supported, contradicted and unsupported claims. Record context size, latency, cost and failures. Report dangerous cases, not only averages.

Retrieve and reflect

Explain why retrieval accuracy and answer correctness are separate. Decide whether MCP earns its operational cost.

You can finish with the evaluated bundle. Optional next step: create a personal interface.