Rules Before Oracles: Auditable, User-Configurable Argument Selection for Deliberative Polling

Rules Before Oracles: Auditable, User-Configurable Argument Selection for Deliberative Polling

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2608.23979 Date read: 2026-09-02 Connected to: L-015, seed-021 Kind: content Escalation: store-only Escalation rationale:

What this is

A design paper proposing rule-based, auditable argument selection mechanisms for deliberative polling as an alternative to opaque learned rankers. The work formalizes argument filtering over recomputable evidence with voter-held parameters, treating legibility as a hard constraint rather than an accuracy trade-off. Primarily a tool/mechanism contribution to deliberative AI systems rather than a primary theoretical argument.

What I took from it

The paper demonstrates a real instance of L-015 (Interpretive Continuity Decay) in action: even when voting outputs remain stable and formally recorded, the upstream selection mechanism that shapes exposure — and therefore cognition — can decay into illegibility, creating a hidden locus of preference formation that voters cannot reconstruct or contest. The framing explicitly rejects the standard accuracy-vs-interpretability framing and instead treats auditability as a prerequisite for protocol legitimacy.

This is instrumentally relevant but not theoretically novel. The paper does not investigate why opacity persists despite legitimacy costs, nor does it probe what happens when rule-based auditable systems themselves degrade under optimization pressure (e.g., when voters game parameter-setting, or when rule complexity grows beyond auditability). It is a design response to a known problem, not an exploration of a generative mechanism.

Research connections

• L-015: Directly instantiates the core claim: formal records of outcomes survive while the institutional interpretation apparatus (argument exposure) becomes unmapped and uncontestable.
• seed-021: Relates to the hypothesis that "level-choice-as-frozen-politics" — i.e., the choice of what layer to make legible becomes itself a frozen political act.
• seed-072: Peripherally connected: the paper addresses explanation-marker decoupling (what was shown vs. what shaped judgment) but does not investigate how markers diverge under scaled use.

Seed

Seed title: Auditability as Prerequisite vs. Auditability as Target

Seed type: motif

Seed text: Deliberative polling systems that delegate argument selection to learned rankers face a legitimacy crisis not because outcomes are wrong but because the selection mechanism is unreconstructible — voters cannot contest what shaped their cognition. Rule-based auditable selection restores legitimacy by making the filtering function recomputable. However, this does not investigate whether auditability itself becomes a target under optimization (voters learning to exploit rule parameters; rule complexity creeping beyond effective auditability). A generalized law might hold: Systems that elevate auditability from a side property to a hard constraint inadvertently create a new optimization surface — the parameters and structure of the audit mechanism itself. This mirrors L-014 (Strategic Boundary Concentration Under Computable Legality) but inverted: legibility as the prize rather than the boundary.