MDCT - Multi-Dimensional Consensus Trees
Introduction - Why do we need MDCT?
Humans naturally organize preferences as hierarchical mental maps. When facing a conflict between working late for a promotion or attending your child's piano recital, you don't just pick one—you weigh entire value systems. You might think "family time matters more than extra income, but if I could reschedule the meeting to tomorrow morning, I could do both." Every decision contains layers of conditionals, fallbacks, and specific requirements.
Communities are simply groups of many individuals trying to reconcile these same complex mental models into collective choices. Yet traditional voting systems flatten this natural why of thinking into reductive either/or decisions that capture only fragments of what people actually want.
MetaPoll's Multi-Dimensional Consensus Trees (MDCT) were designed to mirror the biological preference structures of individual minds and extend that same organizational pattern to groups—creating a collective decision-making tool that preserves the full richness of how humans naturally think about choices. Ending up with collective intelligence at scale.
A bird's eye comparison:
Binary yes/no votes hand control to whoever frames the question—voters can only respond within predefined boxes rather than expressing what they actually want. (Choice while in handcuffs)
Pairwise comparisons explode combinatorially—comparing every security option against every sustainability option becomes intractable. (So costly that they're infeasible)
Multiple choice forces selection of a single option when voters may want a fallback option. (How does a voter express that they like Pizza more than Lasagna but less than Ice cream?)
Single-dimension Ranked choice and Quadratic voting handles preference ordering within one category but cannot express conditional preferences or compare across relationships between categories (How would a voter express if security matters more than cost savings, while inside the cost dimension also express if they prefer to pay more for upfront costs or long term maintenance costs?)
MDCTs organize options into intuitive hierarchies where voters express preferences at any level—from high-level priorities ("performance over cost") down to specific implementations ("Nvidia over AMD")—creating a preference map that captures both what communities want and how different goals relate to each other.
Conditional preferences are a powerful expression feature of MDCT: you might prefer a meat lovers pizza over a burger, but if burger wins, you want a chicken burger with havarti on whole wheat with extra ketchup. People intuitively hold these preference maps in their minds and select their actions accordingly.
Traditional voting cannot capture these nuanced fallback positions and conditional choices that reflect how people actually think about decisions.
What would normally require dozens or even hundreds of separate polls becomes with MDCT a single coherent structure, making preference expression a streamlined activity while giving implementers a unified view of community priorities and tradeoffs.
MDCTs let communities express their natural preference maps—primary choices, acceptable alternatives, and specific requirements within each option—creating rich context for implementation regardless of which path is ultimately chosen.
So why do we need MDCT? Because it translates the full complexity of human decision-making into actionable data—critical infrastructure for organizational intelligence, civilizational coordination, and AI alignment.
The details - Tree Structure
A MDCT organizes options hierarchically. Each option can contain child options, which can contain their own children, forming a tree. Voters can select options at any level of the hierarchy, allowing for both broad category preferences (like "Security" overall) and granular choices (like specifically "24/7 cameras" within surveillance). This flexibility enables complex preference modeling where participants might support a general concept but have specific implementation preferences within that category.
Layer Architecture
Root Layer: Top-level options with no parents. These represent the highest-level categories or priorities.
Layer N: Each level of child options forms a new layer. Layer 2 contains children of root options, Layer 3 contains their children, Layer 4 options contain the children of a Layer 3 options, and so on.
The recursive structure supports arbitrary depth and complexity while maintaining clear relationships across categories. A MetaPoll could theoretically extend 50 layers deep or more if the decision space requires it.
Option Properties
Every option carries three essential properties:
Title: Maximum 80 characters. Brevity forces clarity at each level.
Rank: Integer starting at 1 (highest priority). Represents the option's position within its sibling set.
Rank Weight: Inversely proportional to rank within a layer. In a layer with 6 options:
This weighting system ensures that preference intensity scales with the number of alternatives—more options mean stronger differentiation between top and bottom choices.
Nesting Properties
In addition to the core properties above, options maintain their tree relationships through:
Parent: Reference to the option's parent node (null for root layer options)
Children: Array of references to child options (empty array for leaf nodes)
Dynamic Semantics
A root option like "Security" means nothing in isolation. Its meaning emerges from the ranked preferences of its children. If "Guard staff" and "K9 units" rank highly under Security, the community defines security through active deterrence. If "Cameras" and "Motion sensors" dominate, security meaning shifts to surveillance.
Semantic flexibility based on the ranking distribution of child options allows the same high-level concept to adapt its meaning based on community preferences. The tree structure provides both the framework (what aspects of security exist) and the interpretation (which aspects matter most).
Execution Logic
MDCTs define not just preferences but execution order through strict rules:
1. Depth-first traversal:
Execute the highest-ranked option at the current level. If it has children, descend and repeat. Continue until reaching a leaf node.
2. Mutual exclusion:
When options conflict (e.g., "Open source" vs "Proprietary" under a software project), execute only the higher-ranked option.
3. Compatible execution:
For non-exclusive options, execute in rank order, but only in ways that don't compromise higher-ranked choices. If "Performance" ranks above "Energy efficiency," efficiency improvements cannot reduce performance.
4. Infeasibility exclusion:
Skip options that violate the rules or constraints of the decision space, regardless of rank. Continue checking until finding the highest-ranked feasible option.
All Combined:
Here's how all four rules work together in practice:
Execution logic creates an explicit preference cascade. Every trade-off becomes visible and quantifiable. Teams know exactly what to prioritize and what to sacrifice when resources are limited.
Computational Properties
MDCTs enable several powerful computational features:
Conflict resolution: The structure makes trade-offs explicit and resolvable
Preference aggregation across dimensions: Voters can express preferences at any level of the tree
Automatic priority inheritance: High-level preferences cascade to influence lower levels
Incremental refinement: Communities can add detail where needed without restructuring
The tree structure transforms vague concepts into precise, actionable priorities through collective intelligence. Each layer adds resolution where the community needs it most.
Example: Community Values Cascade
Implementers now have clear guidance: they begin by pursuing the top-ranked value (for example, striving for High Standards under Excellence). If the ideal cannot be fully met, the structure indicates the most appropriate fallback—lean toward the next-ranked option (Satisfactory Completion) rather than oscillating toward lower extremes like Perfectionism or Cutting Corners. In this way, the tree not only prescribes a priority of execution but also encodes practical “if‑then” guidance, ensuring teams know how to proceed and where to pivot when constraints arise.
MDCT and AI Alignment
A fundamental requirement of AI alignment is ensuring humans remain in control—able to effectively communicate what we want to increasingly powerful systems. As AI capabilities grow, they need granular understanding of human preferences to align with our values.
Binary votes and flat preference lists give AI an impoverished view of what communities actually want. Just as LLMs required massive training data to achieve their current capabilities, AI alignment needs rich preference data to understand human values. MDCTs provide this data by capturing not just what we choose, but why we choose it, what trade-offs we accept, and how different values relate to each other.
Each nested decision, conditional preference, and trade-off becomes training signal. When communities use MDCTs, they create machine-readable representations of their collective values that preserve the conditional logic, hierarchical priorities, and nuanced reasoning that define human decision-making. This enables "collective alignment"—AI systems that understand and serve the complex, multifaceted preferences of entire communities rather than reducing human values to oversimplified metrics.
Last updated