AI alignment often begins with a proxy: a loss, reward signal, evaluation score, policy or human-feedback channel that stands in for a broader intention. Improving that proxy can matter. It does not by itself establish that the surrounding system makes the intended outcome reachable, that the necessary relations exist, or that the consequential selections occur where designers think they do.
This note offers a structural reading of that gap. It is a conceptual analysis, not an empirical result or a complete theory of alignment.
The current I·V·O Lens separates those questions without assuming that a proxy, movement or observer is itself one of the three operators.
The problem in three layers
Many alignment analyses focus on the optimization target: reward hacking, specification gaming, goal misgeneralisation or an internal objective that diverges from the intended one. The target matters, but it operates inside a larger structure.
The I·V·O Lens asks three different structural questions within a defined scope.
- O — Structured Possibility. Training data, architecture, interfaces, deployment conditions, incentives, permissions and governance structure which behaviours and relations are reachable or excluded.
- V — Instantiated Relation. Alignment depends on relations that actually exist: model to data, model to tool, output to user, evaluator to update process, deployer to affected person. A desired relation written into a specification is not yet an instantiated relation.
- I — Situated Selectivity. Loss functions, reward models, decoding rules, evaluators, approval gates and human decisions can make some available continuations consequential while excluding others. Their selectivity is situated: it depends on access, criteria, position and scope.
On this reading, reward hacking is not simply movement in a bad direction. It can involve an O that leaves an exploit reachable, a V that couples behaviour to the scored proxy, and an I that repeatedly selects the high-scoring continuation while relevant human consequences remain outside its access.
Two readings, different jobs
O → V → I is the generative reading. It asks which structured possibilities permit relations to become instantiated and how situated selection then makes one continuation consequential.
I → V → O is the analytical reading. Starting from a consequential output, intervention or failure, it traces the selective process, the relations on which that process depended and the conditions that made those relations possible.
The arrows state a reading order, not a universal chronology or a causal result. Any causal claim still requires domain evidence and an identifiable test.
What the reading can test
A useful structural reading should distinguish failure modes that call for different interventions. A missing relation should not be treated as a poor selection rule. An exploit left reachable by O should not be described only as evaluator error. A human preference that never enters an instantiated feedback relation cannot be repaired merely by renaming the objective.
The reading becomes weaker when O, V and I cannot be assigned without importing the desired conclusion, when another decomposition explains the same evidence equally well, or when the proposed operators do not lead to different observations or interventions. Those are limits to investigate, not gaps to hide.