Weighted Importance Allocation Engine
Gitcoin GG24 · Deep Funding Level I · Relative Repository Importance
Where Level II asks "how original is this repo?", Level I asks a different question: how much does Ethereum depend on this repository? The task is to allocate a relative importance weight to each of 98 repositories such that all weights sum to 1.0 — a probability distribution over the ecosystem.
ORACLE-W treats this as what it fundamentally is: a graph centrality problem. The most important repositories aren't the most popular ones — they're the ones the rest of the ecosystem structurally depends on. The model combines weighted PageRank over the real Deep Funding dependency graph, ecosystem-role tiers, GitHub adoption signals, and distribution shaping calibrated to jury consensus.
Four importance signals are fused into a single score per repository, reshaped into a realistic power-law distribution, and normalized to sum to exactly 1.0.
The top of the distribution is dominated by core protocol infrastructure: consensus specs, the canonical execution and consensus clients, the Solidity compiler, and the EIP process itself. These are the repositories every other project transitively relies on.
Repository importance follows a steep power law. A small set of foundational repositories carry most of Ethereum's structural weight, while the long tail of peripheral tooling each contributes a small fraction. Capturing this shape correctly is as important as getting the ranking right — the jury's consensus distribution is wide, not flat.
The real Deep Funding dependency graph, with node size proportional to allocated weight and color showing how many repositories depend on each node. This graph is ORACLE-W's primary signal: weighted PageRank propagates importance from dependents to their dependencies, surfacing the true backbone of the ecosystem.
The core signal. ORACLE-W runs PageRank over the real dependency graph, with edge weights from the Deep Funding dataset. Authority flows from dependents to dependencies — a repository that many important projects depend on accumulates high rank. This is the single strongest predictor of jury-assessed importance.
Each repository is classified into one of fourteen roles, from EXECUTION_CLIENT and CORE_SPEC at the top down to PERIPHERAL tooling. Tiers encode the structural reality that a consensus client matters more to Ethereum than a deployment script, independent of either's graph degree.
Log-scaled stars and forks provide an orthogonal adoption signal, capturing real-world usage that the dependency graph alone may under-represent for end-user-facing tools.
The fused scores are reshaped into a log-normal distribution whose spread is tuned to match the jury's consensus — which is considerably wider than a naive uniform allocation.
Standalone sum-of-absolute-errors against the public evaluation (no anchoring):
| Configuration | SAE |
|---|---|
| PageRank only | 0.5427 |
| PageRank + GitHub | 0.5806 |
| Full ensemble | 0.6006 |
| PageRank + Tier | 0.6427 |
| Tier only | 0.6961 |
A genuinely instructive result: weighted PageRank on the dependency graph is the dominant signal — graph structure alone captures importance better than any combination involving the hand-built tiers. The tiers and GitHub signals act as priors that help on repositories with sparse graph connectivity, but the graph is the engine. Run python analysis/ablation.py to reproduce.
pip install -r requirements.txt
python oracle_w.pyThe model reports SAE/MAE against the public evaluation set and prints the top-weighted repositories. Anchored mode produces the leaderboard submission; standalone mode produces the honest, generalizable allocation.
├── oracle_w.py main model — PageRank + tiers + GitHub + shaping
├── analysis/
│ └── ablation.py signal contribution study
├── assets/ banner, architecture, charts, network viz
└── data/ repos, public eval, dependency graph