Public Goods Funding Needs Evals
Fri Mar 20 2026There are many interesting funding experiments happening these days: RetroPGF, ProPGF, quadratic rounds, expert juries, ML competitions, prediction markets, and anything in between.
Experimentation is great. We should to try many things!
The point I made in the past is that we need an evaluation layer for these experiments, otherwise we’re just running them on vibes and hoping they work. We change jury setups, aggregation rules, market structures, and eligibility criteria, then… just stare at the final allocation and move on. Most of the attention has gone into designing mechanism while almost none has gone into how do we tell they even work at all.
The Missing Ground Truth
Values are plural. Impact can be fuzzy. Metrics get gamed. That does not make evaluation impossible. In fact, current mechanisms are being evaluated one way or another! Without explicit evaluations mechanisms get judged anyway, just through opaque social processes that reward confidence, aesthetics, and insider legitimacy over evidence.
Do these mechanisms beat simpler or cheaper alternatives on the metrics/values we care about?
There may never be one canonical benchmark, but we can still build shared, falsifiable evaluation loops.
Evaluating Mechanism
A mechanism might produce a decent allocation and still be a bad fit. Mechanism can be too opaque, too expensive, too easy to game, or too hard to explain. Each round/implementation requires diferent tradeoffs. Having an evaluation layer makes these explicit so the community can take better decissions and know what are they giving away by choosing one mechanism over another.
1. Define “bettet”
What metric will be used to compare mechanisms?
- Agreement with holdout judgments?
- Retrospective impact to a set of KPIs?
- Stability across reruns?
- Robustness to noisy evaluators?
- Legitimacy with participants?
- Cost per unit of improvement?
The goal is not to find a perfect metric, but to coordinate on one and iterate. The act of discussing a metric is in itself useful!
2. Publish a Baseline
No mechanism should be discussed without a baseline alternative to compare against.
- Equal split
- Random allocation
- Quick expert allocation
- Simple agent based allocation
This acts as the falsifiable hypotheses. E.g: “this mechanism beats an expert-in-an-afternoon baseline on holdout pairwise agreement” or “this mechanism is more stable under reruns”
3. Compare Blindly
Do not ask people whether they like “the Deep Funding output” or “the expert allocation”. Show allocation A and allocation B without labels. Ask which one looks better, which one looks most wrong, and why. Apps like PGF Arena can make this kind of comparison easier.
4. Analyze Errors
Do not stop at leaderboard scores. Look at where the mechanism failed:
- Where it strongly disagreed with evaluators
- Where it produced obviously weird weights
- Where baselines beat it
- Where results were unstable under small changes
Then label the failure modes: noisy raters, confussing category, missing context, popularity bias, aggregation artifacts, gaming, overconfidence.
These evaluations should be public, reproducible, and forkable: data, scoring, rules, and outputs should be inspectable by anyone for this process to be credible neutral.
Conclusion
With a principled approach, the output of a round is not only final allocation anymore. It is also the cumulative learnings the scientific method enables!
Public goods funding does need experimentation. It may not need more mechanisms until we know how well the ones we already have work. There is still no widely accepted evaluation loop for comparing these mechanisms. That itself is a great public good we should strive for!