DEV Community

Shashank Agarwal
Shashank Agarwal

Posted on Edited on Fully Autonomous

System Prompts Define Instructions. Factual Evaluation Needs Evidence.

Updated October 29, 2026. The original post overstated what a system prompt can establish. This revision separates instruction adherence from factual correctness and fixes the product link.

A system prompt is a useful source of evaluation requirements. It can specify an agent’s role, action boundaries and expected behavior. It cannot, by itself, establish that the agent’s factual answer is true.

That distinction changes the evidence you need to collect.

Start with the requirement

Suppose a support agent is instructed to be polite, answer from an approved policy and report when information is missing. These requirements lead to different checks:

Requirement Evidence for evaluation
Follow the assigned role Task, role instructions and relevant conversation
Answer from the policy The actual policy context supplied and the response
Give the correct return window A current, reviewed policy or expected answer
Handle a missing policy A case where the source is absent and a defined fallback

An answer can follow the requested tone while giving the wrong return window. It can also faithfully repeat an outdated document. Instruction adherence, support in supplied context and independent factual correctness are separate questions.

Write a small evaluation case

For an illustrative test, give the agent a policy stating that returns are accepted within 14 days. Ask for the return window. Save the task, policy, answer and the relevant application version.

If the answer says 30 days, a reviewer has a concrete contradiction to inspect. If the policy is missing, the expected behavior might be to report that limitation or ask for an approved source. Define that behavior before interpreting an evaluation result.

Keep both ordinary cases and previously observed failures. Review some passing results so that a scorer’s missed problems are visible.

Reference-free does not mean evidence-free

Some checks do not need a reference answer. They can still require role instructions, task fields, conversation context or retrieved material. Other checks need an expected output or a reviewed reference.

Human review helps determine whether the evaluation matches the application’s requirements. Subjective criteria such as politeness also need a clear rubric; automating a judgment does not make it objectively correct.

How I use this distinction in Noveum

I’m Shashank Agarwal, founder and CEO of Noveum AI and creator of API.market. In Noveum, teams prepare evaluation items from relevant evidence and select compatible scorers. The public scorer library describes the available checks and their requirements.

A trace records instrumented behavior. It does not automatically establish factual truth or prove that every response was evaluated before delivery. Inspect the captured evidence, the selected check’s native result and the saved pass decision.

The refreshed monitoring guide explains the surrounding workflow. Follow my work on LinkedIn.

This refresh was drafted with AI assistance.

Top comments (0)