01
How I got to the question
In Intent-aligned AI systems deplete human agency, Mitelut, Smith, and Vamplew argue that satisfying a person’s current intent may not preserve their longer-term agency. They study agency loss in simple environments and propose further research. I read this as a reason to ask what an agent should preserve besides immediate task success.
My systems work raises a related operational question: when may a retrieved claim guide an action, and when should the system seek another check? These are connected questions, but I haven’t shown that one architecture answers both.
02
What I want to preserve
I want claims to carry their sources, the steps that produced them, and the limits of where they apply. Confidence should not silently increase when a summary is copied into another system. A decision to defer should have a reason that a later reader can inspect.
A stored confidence value is not enough. It needs calibration against outcomes, and a policy connecting uncertainty and consequences to action. I still need to test how much of this helps compared with simpler records and checks.
03
A testable proposal
I would compare workflows with and without explicit lineage and deferral rules, holding the underlying tasks and models fixed. I would look at incorrect actions, unnecessary deferrals, and whether another person could reconstruct the basis of a decision.
This overlaps with provenance, uncertainty, and human-review systems that already exist. My task is to specify the combination I need and show where it helps, rather than claiming nobody has built the missing layer.