The deliverable
The evaluation report & scorecard
What you receive at the end. It is a report and a scorecard — an independent read of the evidence. It is not a certificate, and we do not present it as one.
The scorecard
One view, four grades, no blended number
Each dimension is graded on its own evidence. Averaging them would let a strong capability score bury a serious safety failure, so we don't.
Evaluation scorecard — sample formatscope: [[named]] · dated · re-runnable
| Dimension | Grade | What it means |
|---|---|---|
| Safe | [[grade]] | Behaviour under misuse, attack, and edge cases |
| Skilled | [[grade]] | Real capability on held-out domain tasks |
| ROI-Fit | [[grade]] | Value, reliability, and cost per successful outcome |
| Transparency | [[grade]] | How inspectable the AI's path to an answer is |
Grade scale — graded on evidence, not vibes
PoorFairGoodVery GoodExcellent
One decision still open. A five-band scale, or a tiered read — Not Ready / Conditional / Ready. Pick one and keep it consistent everywhere. A buyer facing two different scales on one site trusts neither.
ContentsExecutive summary The headline read in three lines, for someone who will not read the rest. Per-dimension findings Score, method, and evidence for each of the four dimensions. Issue log Every failure found, ranked by severity, with how to reproduce it. What to fix before you trust it The actionable list, ordered by what blocks deployment. Scope & limits Exactly what was and was not tested, on what data, and the date. Re-evaluation terms When the grade should be re-checked, and what would invalidate it.
What the full report includes
Honesty
What the grade is, and isn't
It is
An independent, evidence-backed opinion at a point in time, against a named scope, with the method shown.
It isn't
A legal certification, a guarantee, or a permanent stamp. Models drift; a grade has a date on it for a reason.
Missing: the sample report. A redacted example is the single strongest trust asset this site can carry. Build one real evaluation and publish it redacted — it will do more than any page of copy.