Designing for Disagreement: Front-End Guardrails for Assistance Allocation in LLM-Enabled Robots

This paper proposes "bounded calibration with contestability," a front-end design pattern for LLM-enabled robots that manages scarce assistance allocation by constraining decisions to approved modes, ensuring real-time legibility, and providing specific contest pathways to address pluralistic values and LLM variability without overburdening users.

The Gist

Imagine you are standing in a busy train station with a friendly robot guide. You need help finding a gate, but at the exact same time, a panicked person drops their wallet and needs immediate assistance. The robot can only talk to one person at a time.

Who does the robot help first?

This is the core problem the paper tackles. With the rise of "smart" robots powered by Large Language Models (LLMs), these machines are getting better at making complex decisions. But they are also unpredictable. Sometimes, a robot might decide to help the person who speaks the loudest; other times, it might help the person who looks most confused. There is no single "right" answer, and different people have different values about what fairness looks like.

The author, Carmen Ng, argues that we can't just let the robot decide silently in the background, nor can we ask every stressed-out user to configure the robot's "moral settings" while they are waiting in line. Instead, we need Front-End Guardrails.

Here is the paper's solution, explained through a simple analogy: The "Traffic Light" System for Robot Help.

1. The Problem: The "Black Box" vs. The "Wild West"

Currently, we have two bad options:

• The Black Box: The robot decides who to help based on hidden code. You don't know why it ignored you, and you can't argue with it. It's like a traffic light that changes color randomly.
• The Wild West: We let users change the robot's rules on the spot. "I want to be first!" "No, I'm more important!" This creates chaos, especially when people are stressed or in a hurry.

2. The Solution: "Bounded Calibration with Contestability"

The paper proposes a middle ground called Bounded Calibration. Think of this as a menu of pre-approved rules that the robot can follow, rather than letting it invent new rules on the fly.

Here are the three ingredients of this system:

A. The "Governance Menu" (Bounded Calibration)

Imagine the robot doesn't have infinite choices. Instead, the station managers (the "governors") have set up a small, safe menu of options, like:

• Option 1: Help the most urgent person first (e.g., someone crying or injured).
• Option 2: Help people in the order they arrived (First-Come, First-Served).
• Option 3: Help the most vulnerable people first (e.g., elderly or disabled).

The robot is bounded because it can only pick from this menu. It cannot suddenly decide to help the person with the shiniest shoes. This prevents the robot from making wild, discriminatory, or harmful choices.

B. The "Signboard" (Legibility)

When the robot has to choose, it doesn't just stay silent. It acts like a traffic cop with a signboard.

• Scenario: The robot helps the person with the lost wallet and tells the tourist, "I am currently in Urgency Mode. I must help the person with the emergency first. I will come back to you in 30 seconds."
• Why it matters: The tourist now understands why they were delayed. They aren't being ignored; they are being queued based on a clear rule. This makes the robot's decision legible (easy to read and understand).

C. The "Appeal Button" (Contestability)

What if the tourist thinks the rule is unfair? Maybe they have a medical emergency too, but the robot didn't see it.

• The Mechanism: The robot gives the tourist a way to say, "Wait, I need to contest this decision."
• The Result: The robot doesn't immediately change the global rule (it doesn't switch from "Urgency" to "First-Come" for everyone). Instead, it triggers a specific path: "I hear your concern. Let me call a human staff member to review your specific situation."
• Why it matters: This gives the user a voice without breaking the system. It's like a "Manager on Duty" button. You can challenge a specific outcome without trying to rewrite the laws of physics.

3. Why This Matters

The paper argues that in a world where robots are making real-time decisions about who gets help, we need to stop treating these decisions as just "computer code."

• It stops "Silent Bias": If the robot picks a default rule (like "help the loudest"), it might accidentally hurt quiet or shy people. By forcing a choice from a pre-approved menu, we make sure the bias is visible and intentional.
• It reduces the burden on users: We don't want stressed travelers to be philosophers debating ethics. We want them to see a clear sign and have a clear way to complain if something goes wrong.
• It builds trust: When people understand why a robot made a choice, and they have a way to fix mistakes, they are more likely to trust the technology.

The Big Picture

Think of this system as a traffic light with a clear schedule and a call button.

• The Menu is the schedule (Red, Yellow, Green).
• The Signboard is the light telling you why you are stopped.
• The Contest Path is the button to call the traffic controller if the light is broken.

By using this "Front-End Guardrail," we ensure that as robots become smarter and more social, they remain fair, understandable, and accountable to the humans they serve.

Technical Summary

Based on the paper "Designing for Disagreement: Front-End Guardrails for Assistance Allocation in LLM-Enabled Robots" by Carmen Ng, here is a detailed technical summary:

1. Problem Statement

The integration of Large Language Models (LLMs) into socially assistive robots introduces complex ethical challenges regarding assistance allocation in real-time, multi-user settings.

• Scarcity and Pluralism: Robots often face situations where multiple users require assistance simultaneously, but the robot can only attend to one. Reasonable people often disagree on prioritization principles (e.g., urgency vs. queue order vs. vulnerability), creating a "pluralistic values" problem.
• LLM Variability: LLM behavior is stochastic and context-dependent. Prioritization policies may vary across prompts and groups in ways that are difficult to anticipate or verify at the moment of interaction.
• The Gap: Current front-end designs lack specific guardrails for these scenarios. Existing approaches tend to fall into two extremes:
  • Silent Defaults: Hidden value commitments that obscure who is being prioritized and why.
  • Open User Configuration: Allowing users to set "value settings" in real-time, which shifts the burden of ethical decision-making onto users under time pressure and risks coercion or conflict.
• Core Issue: There is no established mechanism to make prioritization rules legible, keep them within admissible bounds, and provide a contestable pathway without renegotiating global rules during an interaction.

2. Methodology: Bounded Calibration with Contestability

The paper proposes a procedural front-end design pattern called "Bounded Calibration with Contestability." This is not a specific algorithm but a governance framework for the interaction layer. It consists of three core components:

A. Governance-Approved Menu (Admissibility)

• Concept: Instead of open-ended value settings, the system offers a pre-defined, limited menu of prioritization modes (e.g., "Urgency-First," "Queue-Order," "Vulnerability-Aware").
• Constraint: These modes are defined and approved by human governance (deployers/staff) upstream. Harmful or discriminatory configurations are excluded by default.
• Selection: Authorized roles (e.g., station staff) select the active mode for a specific context window (time/location), preventing individual users from overriding global rules instantly.

B. Continuous Legibility (Abstraction)

• Concept: The active prioritization mode must be communicated to users in interaction-relevant terms at the exact point of deferral (when a user is told to wait).
• Mechanism: The robot explicitly states the rule being applied (e.g., "Priority mode: urgent needs first — I'll return to you next").
• Goal: This transforms the "black box" of LLM decision-making into a transparent communicative act, allowing users to understand why they are being deferred.

C. Outcome-Specific Contest Pathway (Authority)

• Concept: Users who are deferred are provided a lightweight mechanism to challenge the specific outcome without attempting to change the global rule.
• Mechanism: A user can trigger a contest (via voice, button, or operator channel). This triggers a specific recourse pathway (e.g., clarification, escalation to human staff) rather than an immediate mode switch.
• Constraint: The system "role-gates" mode changes (only staff can change the global mode) to prevent responsibility displacement and ensure stability.

3. Key Contributions

• Theoretical Shift: Reframes assistance allocation in LLM-robots from a back-end technical problem to a front-end governance and interaction design problem. It treats value pluralism and LLM uncertainty as standing conditions to be managed, not solved.
• Design Pattern: Introduces "Bounded Calibration with Contestability" as a specific pattern to balance transparency, user agency, and system stability.
• Separation of Concerns: Distinguishes between value mediation (governance-defined rules) and interaction-time allocation (execution), preventing opaque trade-offs during real-time decision-making.
• Scenario Illustration: Provides a concrete vignette of a robot in a busy concourse demonstrating how the pattern handles competing requests (tourist vs. distressed person) through legible deferral and contestation.

4. Results and Evaluation Agenda

Note: The paper is a design proposal and does not report empirical experimental results. Instead, it outlines a rigorous evaluation agenda.

• Proposed Evaluation Benchmarks:
  1. Legibility: Can users correctly identify the active mode and anticipate deferrals?
  2. Procedural Legitimacy: Do users judge the fairness of the allocation based on the process (transparency, contestability) rather than just the outcome (who got help first)?
  3. Actionability: Can users access and complete contest steps under time pressure, and do they understand the consequences?
• Proposed Methods:
  • Vignette Experiments: Comparing silent defaults vs. legible modes vs. legible modes + contestability.
  • Wizard-of-Oz Studies: Stress-testing timing, interruptions, and multi-user dynamics.
  • Governance Workshops: Assessing the feasibility of defining admissible mode menus.
• Risk Assessment: The paper identifies risks such as automation bias (users trusting the robot's stated rule too blindly) and uneven usability of contest channels across different user groups.

5. Significance

• For Deployers: Offers a practical method to bound and disclose prioritization choices without exposing the system to harmful configurations or user-induced chaos.
• For Researchers: Defines new front-end evaluation criteria (legibility, procedural legitimacy) that go beyond traditional back-end LLM performance metrics.
• For Regulators/Auditors: Provides a mechanism to generate interaction-level traces (logs of modes, deferrals, and contests), making ethical risks inspectable and accountable as LLM-robot systems scale.
• Ethical Impact: Moves the field away from "silent defaults" that hide value biases and toward procedural justice, ensuring that even when users disagree, the process of allocation is transparent, bounded, and contestable.