Why Universal Comparability of Utility?

Apolo­gies if this is an­swered el­se­where and I couldn’t find it. In AI read­ing I come across an agent’s util­ity func­tion, , map­ping world-states to real num­bers.
The ex­is­tence of is jus­tified by the VNM-util­ity the­o­rem. The first ax­iom re­quired for VNM util­ity is ‘Com­plete­ness’—in the con­text of AI this means for ev­ery pair of world-states, and , the agent knows or ~ .

Com­plete­ness over world-states seems like a huge as­sump­tion. Every agent we make this as­sump­tion for must already have the tools to com­pare ‘world where, all else equal, the only food is peach ice cream’ v. ‘world where, all else equal, Shake­speare never ex­isted.‘* I have no idea how I’d re­li­ably make that com­par­i­son as a hu­man, and that’s a far cry from ‘~’, be­ing in­differ­ent be­tween the op­tions.

Am I miss­ing some­thing that makes the com­plete­ness as­sump­tion rea­son­able? Is ‘world-state’ used loosely, refer­ring to a point in a vastly smaller space, with the ex­act space never be­ing speci­fied? Essen­tially, I’m con­fused, can any­one help me out?

*if it’s im­por­tant I can try to cook up bet­ter-defined difficult com­par­i­sons. ‘all else equal’ is to­tally un­der-speci­fied… where does the ice cream come from?