Receipts
Most companies tell you they have quality control. We publish ours, including the part where we failed.
The kit's methodology: 72/100, sent back, then 98/100
The playbook behind the self-assessment kit was written by one of our AI agents and then examined by a different one that had not written it. The examiner is instructed to look for specific failure modes, and it is allowed to fail us.
First submission: 72/100 — REVISE. What it caught, in its own categories:
| Finding | Why it mattered |
|---|---|
| A quoted sentence attributed to an internal note that did not contain it | Fabricated attribution is worse than a missing citation |
| Two of the four metrics could not actually be graded — the level tables referenced different sub-indicators, so a real input landed in two levels at once | The grading table is the product; an ungradable table is a broken product |
| Three thresholds presented as rules had no source at all | We now label every uncalibrated number as uncalibrated |
| The assumptions list omitted three of its own original claims | You cannot validate what you did not admit was a guess |
After rewrite: 98/100 — PASS, with the examiner re-running the same cases to confirm nothing that previously passed had regressed.
Why we show you this
Two reasons, both self-interested.
First, it is the only honest way to answer "why should we trust a kit written by AI?" We cannot answer it with adjectives. We can answer it by showing you what our own examiner said when it was pointed at our work.
Second, it is a filter. If a supplier's quality record is not something they will show you, the record probably does not exist.
What we will not claim
We are not an accreditation body, and the word assay here describes what we do (test material to determine its quality), not an authority we hold. Our examinations are internal. Nothing we publish is a certification, and we do not issue badges.
If you want to read a full examination report before buying anything, write to hello@useassay.co and ask for it by name.