Name what you are testing before you pick a tool

6 September 2026

“Test the AI” is not a scope.

If you don't name the object under test, every vendor demo looks like a fit — and none of them will save you when something fails in production.

Here is the short list we use when we walk into a system.

The model

You care about capability, robustness, bias, or safety behaviour of the foundation or fine-tuned model. Benchmarks and scanners help. Remember: once you wrap that model in prompts, tools, retrieval, and filters, behaviour changes. Model scores ≠ application truth.

The prompt or policy layer

You changed the system prompt, the rubric, or the output schema. Hold everything else still and backtest. This is where prompt regression tools earn their keep. They do not magically invent a good oracle for your domain.

The retrieval layer

Search, vectors, hybrid, rerankers, graphs, permissions. The answer is only half the story. You also need document identity, rank, completeness, freshness, and whether paraphrase or a hostile query pulls the wrong chunk — or a chunk the user should never see.

If you only score the final sentence, you will misdiagnose retrieval failures as “hallucinations” forever.

The end-to-end application

User input → routing → retrieval → model → tools → what the user sees. Component metrics and behavioural tests both matter. One blended score usually lies to you.

The agent trajectory

Multi-step agents fail in the path, not only the prose. Tools chosen, arguments passed, approvals skipped, retries, where it stopped. If your harness only captures the last message, you are testing a novel, not an agent.

Production operation

After release: traces, feedback, incidents, drift, cost, latency, and a path from “that was bad” back into a regression case. If your monitor can't preserve enough to reproduce a failure, it is analytics — not an evaluation loop.

A platform that is excellent for one of these can be a poor fit for another. Don't judge a retrieval framework by its prompt versioning. Don't treat an observability product as a security assessment because it stores traces.

Write the object down first. Then pick tools. Next in the series: a map of the landscape by the job to be done — not by marketing category.

Part of breaklight's public notes on how independent AI testing actually works. We still say retrieval, not “RAG shop.”