The judgment layer isn't AI. One logic, rendered three ways — a runnable engine, a spreadsheet, and a live demo — verified to produce the same answer to the byte.
Same inputs→run through all three→same verdict
Engine
runnable Python, zero deps
89
Do not ship
≡
Workbook
the same logic in Excel
89
Do not ship
≡
Demo
live in the browser
89
Do not ship
Byte-identical · same inputs, same verdict, every time
Who does what
The AI
Research assistant
Reads, drafts, synthesizes. Fast and fallible. It's allowed to be brilliant and uncertain.
The engine
The judge
Renders the verdict. Boring and reproducible. Same inputs, same answer — you could hand it to an auditor.
A guess can't be reproduced. A verdict has to be. That's the whole reason the judgment is deterministic and the AI never gets to decide.