Would the answer be the same next time?
The vague word can be defined in advance. The particular fact cannot. That one test decides what a policy is able to hold.
Work that keeps stopping to ask a person a question is not automated, it is just slower. Some of those questions genuinely have to be asked. The useful thing is a test that tells you which is which and there is one.
Would the answer be the same next time?
The questions below came out of a weekly meal plan for a household, which is the example this was built and tested on. The subject does not matter. What matters is which of them a policy could have settled in advance and which ones only this person, this week, can answer.
| the thing it wanted a person for | same next time? | |
|---|---|---|
| what "healthy" means | yes | your standard, not this week's |
| what "not too expensive" means | yes | a budget band is a house rule |
| what "rushed" means in minutes | yes | twenty five minutes is a definition |
| how many children and their ages | no | this household |
| which evening has swimming | no | this week |
The pattern is clean once you see it. The vague word can be defined in advance. The particular fact cannot.
So the person is not removed from the work, they are moved. They answer the words once in the policy and the facts once at the start. Then it runs without them.
What that is worth
Same starting point, same model, one variable.
WITHOUT the definitions BLOCKED at pass 1, nothing adopted, 2 questions handed back WITH the definitions SETTLED in 4 passes, 3 decisions adopted and tagged, 0 handed back
The claim that supports is not the one it looks like. It is tempting to read it as the model behaving better when it can see where things came from. It does not. Recording where a value came from changes nothing about the behaviour. Decisions made in advance change the behaviour. What the record does is let you see afterwards which decisions those were.
The near miss, which is the part worth keeping
A run with definitions in place can settle in two passes with no questions and look like a better result than the one above.
It is not. Apply the definitions additively and the object ends up holding both the vague word and its own definition, plus a default sitting next to a sharper value that was already there. Same headline number, quietly worse.
"Settled with no questions" is not the measurement. What is in the object at the end is.