Summary: Deep Funding is best understood not as a generic AI grants experiment or another grants frontend, but as a mechanism that turns public-goods allocation into an edge-weighting problem over a dependency graph. Instead of asking reviewers to rank whole projects directly, it asks model builders to assign how much credit should flow from a dependent repository to each of its dependencies, then uses human spot-checkers to evaluate a sampled subset of those judgments. The reusable mechanism insight is that practical allocation power moves upstream into graph construction, seed-node selection, edge-definition policy, jury question design, and model-scoring rules. That makes Deep Funding a useful comparison class for Drips, Open Source Observer, Pairwise, and Project Catalyst: it sits between measurement infrastructure, human judgment, and final capital distribution rather than cleanly belonging to only one of those layers.
What it does:
Builds a directed dependency graph for an ecosystem and asks participants to submit edge weights that represent how much credit should flow from a source repository to its dependencies
Uses open model submissions — not necessarily only conventional ML models, but also heuristic or hybrid allocators — to propose those weights at graph scale
Uses a human jury to spot-check sampled comparisons instead of manually reviewing the full graph, then scores submissions by compatibility with those human judgments
Applies the best-performing submission or submissions to distribute sponsor funds across upstream open-source projects
Publishes graph data, scoring code, and related infrastructure openly through the deepfunding GitHub organization
Decomposes the challenge into layers such as seed-node comparison, seed-versus-child retention, and child-node dependency weighting rather than treating project impact as one flat score
Key claims:
The official site states the mechanism plainly: Deep Funding uses “a market of AIs as the engine and humans as the steering wheel.” That is analytically useful because it frames the project as distilled-human-judgment infrastructure, not just automated ranking.
The deepest reusable primitive is the edge-weighting rule itself. In the dependency-graph repo, a model does not submit one score per repo; it submits weights on edges where source -> target means the dependent repo attributes some portion of its credit to that dependency, while the leftover weight remains with the source. That is a cleaner and more composable mechanism than generic project scoring.
The official site’s three-level challenge structure is especially worth keeping. Level 1 compares seed nodes against Ethereum, Level 2 decides how much credit stays with seed nodes versus passes downstream, and Level 3 assigns weights across thousands of child-node dependencies. This makes Deep Funding useful as a multi-layer credit-routing system, not merely a one-shot leaderboard.
The scoring repo is important because it shows the mechanism can operate in more than one evaluation geometry. The project explicitly distinguishes a single-layer mechanism from a two-layer / three-layer mechanism, which helps separate jury design from graph design instead of flattening everything into AI funding rhetoric.
Public materials also reveal where the real control surface sits. The official website, Gitcoin explainers, and the current GitHub repo disagree on graph size and challenge parameters: the site describes 34 seed nodes, over 5,000 child nodes, roughly 15,000 weights, and a 2025 challenge window; Gitcoin pages describe roughly 40,000 Ethereum dependency edges and an initial $250,000 sponsorship; the current dependency-graph README describes 31 seed nodes, 5,024 dependency nodes, and 14,927 edges after graph updates. That version drift is analytically valuable because it shows that graph-construction policy is itself governance.
The model competition is only half the story. The human spot-check layer determines what counts as alignment, and the Deep Funding materials repeatedly emphasize that jurors answer comparative questions rather than writing absolute allocations. So the practical authority is split between model builders, graph maintainers, and jury-interface designers.
Pairwise’s official Deep Funding post makes that human layer more concrete: jurors compare repositories side by side and those comparisons become the reference point for selecting the model closest to human judgment. This matters because it ties Deep Funding to ballot-construction middleware and shows that seemingly AI-native allocation still depends on upstream human-comparison UX.
Deep Funding belongs in the active corpus because it gives the grants and public-goods cluster a stronger dependency-directed comparison point. Drips shows developer-configured upstream splits, Open Source Observer provides measurement inputs, Pairwise provides human comparison tooling, and Deep Funding turns those ingredients into a mechanism for selecting graph-wide credit routes at ecosystem scale.
Whitepaper: No canonical standalone PDF surfaced in this pass. The strongest materials were the official site, the public mechanism draft Deep Funding: A Prediction Market For Open Source Dependencies, the GitHub dependency-graph and scoring repositories, Gitcoin’s public explainers, and Pairwise’s official Deep Funding post. See ../whitepapers/deep-funding-primary-sources-2026-05-11.md.