Skip to main content

Specs

An eval scores outputs against a known-good answer. But often you don't have the known-good answer yet — that's the hard part. A spec is how you get it: instead of scoring what the AI produced, you gather what a correct answer should be from the people who'd know, find consensus, and turn that into the golden answer an eval can then score against.

How it works

  1. Pick the cases

    A spec runs over a dataset of input-only Q&A items — the questions you want ideal answers for.

  2. Gather responses

    Contributors write their own ideal answer for each item, in plain language. Invite teammates, or send a share link so domain experts can contribute without an account.

  3. Find consensus and promote

    Compare the responses side by side. When the right answer is clear, promote it — it becomes that item's expectedOutput. The dataset is now eval-ready.

A spec is IN_PROGRESS while you collect, CLOSED once you've promoted what you need.

When to reach for one

  • You're standing up a new eval and have questions but no answer key.
  • The "right" answer is a judgment call that belongs to experts: support leads, lawyers, clinicians.
  • You want the answer key to reflect what several people agree on.

Spec presets give you a starting frame for common shapes — support responses, code review, policy interpretation, tone & voice, advisory framing — so contributors know what kind of answer you're after.

The loop it closes

Spec → promote → expectedOutputeval. You gather the ideal once, with people, then score against it automatically forever. That's the bridge between "we think it's good" and "we can prove it."

2 min read