Pairwise allocation leverages relative preferences to streamline the cognitively taxing process of allocating resources among many projects. Participants compare pairs of projects (e.g., “A is better than B”) rather than evaluating all projects simultaneously or assigning absolute scores. These pairwise comparisons are aggregated into a graph structure where preferences “flow” through the network until reaching a steady-state global ranking. This global ranking determines how funds are distributed among the projects. Unlike traditional allocation mechanisms, pairwise preference simplifies decision-making and enables more granular and nuanced evaluations, especially for large communities.

First proposed in 2018 by Daniel Kronovet, Aron Fischer, and Jack du Rose from Colony. Pairwise was initially conceptualized as a variation of Google’s PageRank algorithm adapted for capital allocation. In 2023-2024, General Magic implemented the model for Optimism’s Retroactive Public Goods Funding (RetroPGF) rounds, iterating its design over several cycles.

Advantages

• Cognitive Simplicity: Reduces mental load by breaking down complex allocation decisions into simple binary choices.
• Preference Discovery: Helps surface high-quality but lesser-known projects by enabling direct comparisons regardless of reputation.
• Signal Quality: Produces more nuanced distributions by focusing on relative value rather than absolute scores

Limitations & Risks

• Conversion Challenge: Translating relative preferences into absolute allocations can feel opaque and reduce participant trust in outcomes.
• Decision Volume: The minimum number of required comparisons for the outcome to represent global preference is not always clear.

Design Considerations

• Evaluators Assignment: Balance workload and result accuracy. Options include random sampling, assigning evaluators a subset of choices, reputation-based weighting, prioritizing experienced evaluators, and quota-based assignments, capping evaluations per participant.
• Tiered Sorting: Reduce decision volume while preserving accuracy. Consider pre-filtering, where low-signal projects are grouped separately before detailed comparisons, progressive ranking, allowing evaluators to refine rankings in multiple passes, and confidence thresholds, skipping redundant comparisons when rankings stabilize.
• Information-Maximizing Pairs: Design adaptive pairing algorithms that prioritize showing project combinations that will yield the most valuable preference data.