Daisy Opens Up — and the Dead End We Found Right After
By DigiEntropy Engineering · 2026-09-06 · 12 min read
We spent a day removing the five walls between Daisy and the people using the site. The next day a request dead-ended on a plan card pointing at a form that was never rendered — and our eval suite had scored that turn a perfect 1.00.
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 AI co-pilot inside DigiEntropy. She has 35 real engines behind her — cold-room loads, compressor selection, envelope checks, refrigerant ranking — and a rule against inventing a number she did not compute. On our eval suite she has been scoring in the 0.9s for months.
She was also, until this week, almost impossible for a customer to talk to.
This post covers two days of work that belong together. The first day removed five walls between a strong compute core and an actual user. The second day paid for it: with the door open, someone asked Daisy to design a cold room, and got a plan whose third step said "confirm the prefilled inputs below" — with nothing below it. The conversation dead-ended pointing at a card that did not exist.
The interesting part is not the bug. It is that our eval suite scored that exact turn 1.00.
35 - real engines behind Daisy - and a rule against inventing a number
Daisy is the AI co-pilot inside DigiEntropy. She has 35 real engines behind her — cold-room loads, compressor selection, envelope checks, refrigerant ranking — and a rule against inventing a number she did not compute.
23,000 - prompt tokens per turn - a guest allowance is about five turns
Guests now get Daisy too, metered per IP against a daily token budget, on a cheaper model. One turn costs roughly 23,000 prompt tokens, so a guest allowance is about five turns.
908 - documents Daisy can now search - 803 of them blog sections from 89 posts
What Daisy can now search: 908 documents, 803 of them blog sections chunked out of 89 posts, plus the site directory, brand pages, application guides and the curated notes.
55 - pages with adapters - against 24 without
Eleven read-only adapters closed that: the catalog now reports 55 pages with adapters against 24 without.
1.00 - eval score of the dead-end turn - now 0.75 under the always-on rule
The broken turn contained a plan. It scored 1.00. On the incident turn the always-on rule drops the deterministic score from 1.00 to 0.75.
0.14 % - false-fire rate on replay - 4 of 2,761 saved eval turns
…