The evaluation maturity model

Every organisation deploying AI is somewhere on the same ladder. Knowing which rung you're on is more useful than knowing what the top looks like, because the next rung is usually the only one you can reach this quarter.

Stage 1 — Vibes

Someone tries the system, likes it, and says so in a meeting. Decisions get made from demos and anecdotes.

This isn't stupid. At the very beginning it's the only available instrument, and it's fast. The problem is that it doesn't degrade gracefully: as usage grows, the anecdotes keep arriving at the same rate while the number of decisions they're being asked to support grows enormously.

You're here if: the strongest evidence anyone can produce is "the team likes it."

Next rung: write down twenty cases with correct answers. That's it. Twenty.

Stage 2 — A spreadsheet

Someone has a set of test cases in a sheet. They run them manually before big releases and record pass/fail by eye.

This is a genuine improvement and it's where most companies actually are. Two things are still missing: the grading standard lives in one person's head, and the cases were chosen because they were easy to write.

You're here if: you have test cases but no written rubric, and one person does the grading.

Next rung: write the rubric. Specific criteria a second person could apply and reach the same verdict. Then have a second person apply it and see whether they do.

Stage 3 — An automated suite

The cases are in code. Grading is automated. It runs on a schedule and produces a number.

This feels like arrival, and it's where the dangerous illusion lives. You now have a confident, precise number produced by a process nobody has audited. If the rubric is ambiguous, the automation makes it ambiguous at scale and with a decimal point.

Three questions that separate a real stage 3 from a cosmetic one:

• Has anyone checked that the known-correct answer scores full marks against your own grader?
• Has anyone tried to beat the grader with a deliberately wrong answer?
• Do you know your run-to-run variance?

If the answer to all three is no — and it usually is — you have automation, not measurement.

Next rung: the two checks. Oracle scores 100%. Cheats score 0%.

Stage 4 — Calibrated

Humans grade a small held-out set. The automated grader has demonstrated it agrees with them, above a threshold, on cases it hadn't seen. Model versions and environments are pinned. Results carry intervals.

This is the first stage where the number can be defended to someone who wasn't in the room and didn't want to believe it.

The cost is real: you need expert hours, and you need to accept that some things you were measuring turn out not to be measurable as written. Teams passing from 3 to 4 almost always discover that a chunk of their existing suite was broken.

You're here if: you can state your grader's agreement with human experts, with a sample size, and you gate on it.

Next rung: make the record external — signed, dated, and checkable by someone else.

Stage 5 — Verifiable

The result is signed and can be verified by a third party without contacting you. The audit trail includes what was refused, not only what passed. Provenance and licensing are attached to every task. The method is published in enough detail to be rebuilt.

Most organisations don't need stage 5 for internal decisions. You need it the moment a result has to travel — to a regulator, a board, an acquirer, or a customer's procurement team who has no reason to trust you.

You're here if: someone who dislikes you could check your number and would get the same one.

The honest diagnostic

Two questions place you accurately, and both are uncomfortable:

Could a competent sceptic reproduce your headline number? Not agree with it — reproduce it. Same tasks, same grader, same result.

What's the last thing your evaluation stopped you from shipping? If the answer is "nothing," you don't have a gate. You have a report that happens to be numeric.

Why skipping rungs doesn't work

The temptation at stage 2 is to buy a platform and land at stage 4. It doesn't work, for a specific reason: stages 3 and 4 both require a rubric that survives two people applying it, and no tool writes that for you. The tool can enforce it, version it, and calibrate against it. It cannot decide what "correct" means at your company.

That decision costs a few hours of your best practitioner's time and it is the only irreducible cost in the whole discipline.

Everything else is engineering.