Trust model

How the radar ranks.

RepoRadar is an evidence-linked AI discovery radar. It ranks useful AI tools, repos, models, papers, and releases while keeping popularity, evidence, risk, and editorial verdicts visible and separate.

How items are found

RepoRadar monitors official APIs (GitHub, Hugging Face, arXiv, PyPI) and public launch / discussion sources (Hacker News, blog changelogs). Items enter the radar as raw signals; collectors normalize, deduplicate, and score them.

How scores are calculated

The Practical Value Score is a formula that weights:

  • 40% — Usefulness (does it solve a real problem?)
  • 18% — Maturity (stable, documented, public)
  • 14% — Maintenance (active commits, recent releases, open issues)
  • 10% — Documentation (README, examples, model card)
  • 13% — Novelty (recent, distinctive)
  • 5% — Safety / trust (license, security policy, code provenance)

Risk is separate

Risk is not hidden inside the value score. Each card shows its own Operational Risk label (None / Conditional / Medium / High) with a plain-English basis. A repo can be highly useful and high-risk at the same time.

Confidence

Confidence is a 0.0 – 1.0 score plus a label (Low / Medium / High) based on the strength of evidence: README, official API, releases, issues, model card, and review depth.

Tiers

  • Gold — top 10% of eligible items or score ≥ 7.5 with Medium+ confidence and clear public evidence.
  • Silver — score ≥ 6.5, useful but not exceptional.
  • Bronze — score ≥ 5.25, useful or niche, may need more evaluation.
  • Low Signal — score < 5.25, weak evidence, or low transparency.

What does not affect ranking

  • Paid placement, sponsorship, or affiliate deals.
  • Personal relationships, conflicts of interest.
  • Hype, social media popularity alone, or celebrity endorsement.

What does affect ranking

  • Public evidence (docs, code, releases, issues).
  • Activity and maintenance signals.
  • Adoption and momentum, when measurable.
  • Editorial review depth.

Known limitations

Hands-on testing is rare unless explicitly marked. Most cards reflect automated MVP triage plus a documentation / release / issue review pass. The radar is a launch product; the scoring and risk labels will be calibrated against real duplicate and miss logs as the catalog grows.