Part one of three
The doctrine
Can I trust our name to this work?
A decision-maker asks that before the company's name goes on work an AI agent produced. A practitioner asks the same of their own name.
Evaluator Driven exists so the honest and confident answer can be yes.
Who it is for
The doctrine names three readers.
- The decision-maker, who sanctions handing a kind of work to AI and must put the company's name on the result.
- The practitioner, who already uses AI and must put their own name on what it helped make.
- The practitioner's AI agent, which needs a published standard to work to, not its own improvisation.
Doctrine §1.
How it works, in four lines
- Set the bar before the work begins.
- The work is made against it.
- Someone who did not make it judges it.
- The record stays with the work.
Doctrine §5, "Four moves".
The bet
The same job, run inside a well-managed and well-judged loop, beats the same job run without one.
This is a bet, and the doctrine calls it one. It is not proven.
What it is not
- Not a method for making things. Make whatever products you like with whatever tools you like: by specification, by proof, by hand, by agent.
- Not a product. It is a set of rules you adopt, not a product you buy.
- Not certification. A record shows how a judgement was made, not that the work is correct.
Doctrine §9, three of its bullets, quoted.
Status
Version 0.1, ratified 2026-10-25. Licence: CC BY 4.0.
The other two parts
- Custody record — the record that work carries, so anyone can check how it was judged.
- Implementations — tools that run the loop and keep the record.
The doctrine never redefines the record's terms, and the record never depends on the doctrine. A job run under another method can still produce a valid record.
Doctrine §7, quoted.