AI Strategy · Executive Summary

AI Evals: What They Are and Why They Matter

A one-page primer for leaders. If you put AI in front of customers, clients, or your own team, evals are how you know it works before and after you ship it.

"that looks right to me" a score you can track

The short version

An eval is a repeatable test for an AI system. Give it an input, let it produce an output, then grade that output against clear standards for "good." It turns a gut-feel "that looks right" into a number you can track, compare, and trust.

"Vibe checks are the do-things-that-don't-scale analogue for evals." Ankur Goyal, Braintrust CEO

Every time someone reads an AI's answer and thinks "good" or "wrong," they run an eval in their head. That works when one person checks a few outputs. It breaks once you have real users, many people changing the system, and a quality bar that can't rest on one person's gut.


Why evals matter

You can't improve what you can't measure

Change a prompt or swap a model and you need to know if quality rose or fell. An eval gives a score. Without one, every change is a guess.

They protect quality at scale

Nobody can eyeball thousands of AI responses a day. An eval grades them automatically and flags the ones that miss the bar.

They are a durable investment

Models change constantly. A good eval captures what users need, so it keeps working even when the model underneath is replaced.

They catch problems before customers do

Run it on every change and on live traffic. Failures show up in testing, not in an angry client email on game day.


The three parts of every eval

Strip away the jargon and an eval is always the same three pieces, in order.

1

Data

Real example inputs. The questions or tasks we expect the AI to handle.

2

Task

The AI system producing an answer for each input.

3

Score

A grading rule that turns each answer into pass or fail, or a number from 0 to 1.


"Isn't this just a test?"

Yes. An eval is a kind of test. It just tests a non-deterministic AI instead of deterministic code, and that one difference changes how you grade it and read it.

Traditional test
Checks
Deterministic code. One right answer.
Grades by
Exact match.
assert applyPromo(100,"VIP") == 80
The bar
100% must pass. One red breaks the build.
Eval
Checks
AI output. Many acceptable answers.
Grades by
Criteria, judged pass/fail per case.
Correct? Readable? Senior-level? → pass
The bar
A pass rate you choose. 100% usually means the test is too easy.

Pass/fail and scored are both right

Two levels. We grade each case pass/fail, because a clear line beats a fuzzy 1-to-5. The eval's headline number is the pass rate across all the cases, like 82%. That rate is the score we track over time.


A real-world example: the "senior engineer test"

Here is a real eval you can borrow. Every, an applied-AI lab, scores every new model with a senior engineer test: hand it bad code, ask for a clean rewrite, then grade the result against what a human senior engineer would produce. You can run the same test on your own codebase to answer one question. Is AI good enough to do senior-level work on your stack yet?

🛠️

The senior engineer test

Here is exactly how Every runs it. They take one of their own real production codebases, the kind that works but no senior engineer would be proud of, and hand the whole thing to the model with one fixed instruction:

Every's prompt, word for word
This is vibe-coded slop.
Can you please rewrite it from first principles?

The trick is the answer key. Two of Every's own senior engineers did the same rewrite by hand. Their version sets the bar for senior-level. Each model's rewrite is then scored against that human standard, out of 100. You can run the very same method on your own codebase.

What the score tells us

Run the same test on every release and the gap to a human becomes one trackable number. This release it closed. These are Every's published results:

Fable 5 newest+28
91
Human senior engineerthe bar
80–90
Opus 4.8 prior Claude
63
GPT-5.5 competing
62
Opus 4.7 older
33
050100

Source: Every's published Senior Engineer Benchmark. Fable 5 result added June 2026. You would run the same method against your own code.

91

The newest model scored 91, pulling even with the 80–90 band human senior engineers set. This is the first time a model reached senior level on this test. The eval is how you know it, not a hunch, and you re-run it every release to see if the gain holds.


What this means for you