Choosing the Lens

A team proposes an architectural change to a production service. The proposal collects arguments about latency, operational complexity, deployment cost, and regression risk. Nobody disputes these claims as facts. What is disputed is how they bear on the decision. A performance reviewer reads everything through latency. A reliability reviewer reads the same arguments through incident risk and is unmoved by latency gains without stability evidence. The finance reviewer cares about cost. The arguments and their conflicts are fixed. Which attacks succeed depends on which lens governs the review.

The sponsor of the proposal has some influence over the lens. They choose which review tracks to pursue, who joins the early design discussions, which forum hears the proposal first. They cannot change how a reviewer ranks considerations within a lens. The reliability reviewer will always put stability first. But they can decide which lenses are in the room at all. That is a strategic action, and standard argumentation theory has no parameter for it.

In Dung’s classical model, you have arguments, an attack relation between them, and semantics that tell you which sets of arguments are jointly acceptable. Value-based argumentation gets closer to the reviewer story. Each argument promotes a value, and an audience is a ranking of the values. An attack fails when the attacked argument promotes a value the audience ranks strictly higher. Different audiences accept different arguments from the same structure. But an audience is a strict total order on the values. It can rerank them. It cannot switch one off.

We propose context-dependent argumentation frameworks, where a defeat function decides, per context, which attacks succeed. The specialisation we study assigns each argument a source perspective. A context activates a subset of perspectives and ranks them by priority. An attack succeeds when the attacker’s perspective is active and ranks at least as high as the target’s. There is a deliberate asymmetry here. Deactivating a perspective silences the attacks its arguments mount, but the arguments themselves remain present. They can still be attacked, still be defended, and still be accepted. The perspective loses its voice as an attacker, not its standing. We read the activation as the agent’s action space and the priority as institutional structure the agent cannot touch.

The paper works through a small example. The target argument t and its neighbour a share a perspective, and a attacks t. The only argument that can defend t against a is b. But b also attacks t. So whenever b is ranked high enough to defeat a, it is also high enough to defeat t. The defender doubles as an attacker. Under full activation this trap holds for every injective priority, and t is rejected. The way out is to deactivate the perspective of a and t. The friendly fire from a disappears. b still attacks t, but a third argument d, from a perspective that stays active, defeats b and defends t. So t is accepted while still under active attack. This is the part I find interesting. Deactivation did not simply remove the attackers of t. It rearranged the structure so that t could be defended.

The same example separates the formalism from value-based argumentation. Another winning move for the agent is to activate only the perspective of d. Then both directions of the mutual attack between a and b are inactive. No audience can produce that pattern. A strict ranking always lets at least one direction of a mutual attack succeed. So perspective activation offers strictly more strategic options than audience choice.

The strategic question generalises to a decision problem we call activation manipulation. Given the framework, a fixed priority, and a target argument, is there a nonempty activation under which the target is accepted? The activation must be nonempty, since an empty one deactivates every attack and accepts everything, which we read as the absence of a regime rather than a lens. The problem is NP-complete for stable and preferred semantics. The preferred case looks worse at first, because verifying a preferred extension involves a maximality check. But credulous preferred acceptance coincides with credulous admissibility, so the activation and an admissible witness can be guessed together in one NP computation. For grounded semantics we only have membership in NP and hardness for P. Whether the freedom to choose the activation makes the grounded variant hard is open.

This is a short paper built around one worked example and baseline bounds, so the list of open ends is long. Activations could be constrained or carry costs. The priority could itself be a choice variable. The version I find most promising is the multi-agent one, where several agents jointly determine the activation under cooperative or adversarial dynamics, and one can ask about equilibria. One applied setting where I expect the phenomenon to matter is multi-perspective agent memory, where the same experience receives goal-conditioned encodings reconciled by argumentation at retrieval. Whether it arises there at scale is open too.

The paper has been accepted to the LAMAS&SR workshop at FLoC 2026. Full paper is available here.