Daisy Is Not a Chatbot: The Architecture of an Engineering Co-pilot
By DigiEntropy Engineering · 2026-09-13 · 25 min read
Daisy does not sit beside DigiEntropy as a chat window. PageUx gives her a semantic nervous system, real engineering engines give her instruments, tiered context gives her attention, and three separate evaluation loops keep the whole thing honest.
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.
Most AI assistants live in a rectangle. You type into the rectangle; a model predicts some words; the words stay in the rectangle.
Daisy lives in the product.
She can see which compressor, refrigerant, operating point and result table you are looking at. She can ask a real selection engine to calculate, drive a declared control on the page, explain the numbers already on screen, carry a goal across several tools, and stop at a human confirmation gate before anything leaves the system. The chat rail is only her voice. The product itself is her body.
35 - declared engineering and connection tools - each with a typed input contract
Specialisation lives in the instruments around it: 35 declared engineering and connection tools, each with a typed input contract and a real executor behind it.
79 - pages mapped by PageUx - 1,047 typed elements, 468 declared actions
The current PageUx footprint, measured from the generated catalog: 79 mapped pages, 1,047 typed elements, 468 declared actions and 196 guided workflows.
0.14% - dead-end guard fire rate - 4 of 2,761 saved evaluation turns
Replaying it over 2,761 saved evaluation turns fired 4 times — 0.14% — and all four were genuine dead ends.
79 of 87 - turns left the user something clickable - up from 28 of 87
When we first measured it, only 28 of 87 evaluation turns gave the user anything clickable. After the rule, 79 of 87 did, with 199 routes emitted and none of them invalid.
44.4% - cut in prompt tokens - tiered context, 42.0% in cost
That is a 44.4% cut in prompt tokens and 42.0% in cost — and a small quality dip with two extra gate failures that we did not paint over.
0.004 - quality gap, budget vs premium model - one fifth of the noise floor
By September, on the same architecture but a far better harness, the same two models sit at 0.933 and 0.929 on 110 scenarios. The cheaper one is marginally ahead. The difference is 0.004, which is one fifth of the noise floor.
12.4× - the premium model's bill per turn - for the same work
Median model cost per turn was 0.0069 US dollars on the budget model against 0.0853 on the premium one — the same work, 12.4 times the bill.
…