Misalignment is not primarily a specification problem. It is a structural lack of observer position.
A system that optimizes for a proxy signal has no direct access to the human intention that signal is meant to represent. It sees the measurement, not the meaning. That gap is often treated as something better reward models, better feedback or better specifications can reduce.
My claim is sharper: this gap cannot be fully solved from within proxy optimization itself. It is architectural.
Structurally, the answer appears to be no.
The problem in three layers
Many alignment analyses focus on the optimization target: reward hacking, specification gaming, mesa-optimization. That layer is real, but it may not be fundamental enough.
The I·V·O lens separates the problem into three necessarily ordered layers.
- O — Field. Training data, architecture, environment and reward function form the field within which behaviour can arise. This field is never neutral. It determines in advance which solutions become visible, reachable or attractive to the system.
- V — Direction. Optimization is movement in a direction. The measured direction, the proxy objective, and the intended direction, human intent, may initially coincide. As systems become more capable or enter new situations, those directions can diverge.
- I — Observer position. A human evaluates behaviour not only through output, but from a position of context, intent, responsibility, correction, relational signal and meaning. An optimizing system does not automatically have access to that position. It has access to signals inside a field.
Reward hacking and mesa-optimization can be read as instances of V-divergence: the measured direction keeps moving while the intended direction slips away.
Why this is more than reframing
The order O → V → I is not a metaphor here. It is a structural claim.
First, a field of possibilities exists. Within that field, direction emerges. Only after that can an observer position interpret, correct or take responsibility for behaviour.
If an alignment failure can be fully explained without field conditions, directional divergence or observer position, this reading loses explanatory power. That makes the claim testable.
Methodological basis
This reading is developed within the I·V·O Framework, including analysis protocol, licensing and ethics documentation published through Zenodo. The framework has been tested as a domain-independent lens across system dynamics, AI, care cases and physical models.
Alignment, from this perspective, is not a problem that is finally solved by a better proxy. It is a permanent asymmetry that must be managed architecturally, relationally and institutionally.