Media
3 items
Description
Exposes fake stars on GitHub repositories using statistical anomaly detection. real-stars analyzes the star history of any GitHub repo you visit and shows how many of those stars look real, right next to GitHub's official star count. How it works — two independent fake-star detection signals: • Burst detection: a sliding-window MAD (median absolute deviation) algorithm spots statistically anomalous spikes in the star timeline — the fingerprint of "bought a batch in one day". • Per-user account analysis: samples 200 stargazers and scores account age, follower count, public repo count, and avatar to spot throwaway / farm accounts — catches "drip-fed bought stars from a pool of 6,000 empty profiles" that burst detection misses. • Cross-validation: real spikes leave evidence in fork activity and traffic referrers (HN, Reddit, Twitter); bought stars don't. Sign-in: one-click GitHub OAuth, no PAT setup needed. Calibration: Validated against StarScout's published ground truth (ICSE 2026 paper, https://arxiv.org/abs/2412.13459). Live algorithm matches StarScout's snapshot numbers within ±3% on the test set: - LupusLeaks/EasyFN: 86.5% fake (StarScout: 83.5%) - microsoft/vscode: 1.5% fake (StarScout: 1.27%) - torvalds/linux: 1.5% fake (StarScout: 0.88%) Repos under 1,000 stars get a "needs more data" badge instead of a verdict — we'd rather under-detect than libel a real project. Source: https://github.com/serenakeyitan/real-stars Privacy: real-stars only reads public repo metadata via the GitHub API using your own OAuth token. Nothing is sent to any third-party server except a small Cloudflare Worker that exchanges the OAuth code for a token (the token never leaves your browser after that).