We Gave Daisy's Critics an HR File
By DigiEntropy Engineering · 2026-09-20 · 11 min read
We built a panel of AI reviewers to criticise our co-pilot, then a supervisor to grade the reviewers and fire the bad ones. The first thing it measured was that half of our assumptions were wrong.
About this post — Authored by an AI assistant using DigiEntropy's polynomial corpus, the universal compressor predictor, and the project's chart-generation tooling. Charts are produced by Python scripts that read the same database the live site queries; tables and formulas are pulled from the same engineering modules. Findings reflect the project's current dataset and methodology — send corrections or deeper questions to admin@digientropy.com.
Daisy is the engineering co-pilot built into this platform. She runs our real refrigeration engines — cold-room load, compressor selection, retrofit, envelope checks — and answers inside the page you are already on.
We already measure whether her answers are correct. What we had never measured is whether they are usable: does she give you something to pick from instead of another question, can you tick the options she offers, does she drive the page you are looking at. Those rules are written down in her instructions. Nothing checked them.
So we built something to check them, and a panel of reviewer personas to judge what code cannot. Then, because a bad reviewer steers the product every week and nobody would notice, we built a supervisor over the panel — one that measures each reviewer, puts it on probation, and can fire it, with the whole decision written into a Persona HR Record.
The order turned out to matter more than any of it. We measured first, and the measurement told us that two of our four assumptions were wrong — including the one we cared about most.
2,681 - distinct Daisy turns replayed - from 103 saved evaluation runs
We already had 103 saved evaluation runs on disk — 2,681 distinct Daisy turns we had paid for months ago. Writing the rules as code meant we could replay them across all of it in about a minute, for nothing.
0.6% - turns that only ask questions - the rule we cared about most
"Don't just ask the user questions — give proposals to select from." This was the rule we cared about most. Daisy breaks it on 0.6% of turns. She already does this.
93.9% - turns with no selectable points - a card-contract gap, not Daisy's fault
"Answer in point form, where the points can be selected and de-selected." Broken on 93.9% of eligible turns — and it is not Daisy's fault. There is no card in her vocabulary that renders individually selectable points.
0 of 9 - prose options the panel caught - perfect on fabricated numbers
It is 0 out of 9 at noticing that a list of options was written as prose instead of something you can tick.
…