Screen
AI-assisted screening surfaces projects with measurable momentum across GitHub and Hugging Face—not merely loud launch-day claims.
Runeval methodology
The signal is only the start.
Popularity tells us where to look. Careful research, source verification, and a final quality check determine what deserves to be published.
SignalEvidenceQuality
Our process
Three steps between momentum and publication.
Research
Technical documentation and relevant web research inform each analysis. Material claims are traced, deductions are explained, and conflicting evidence is preserved.
Quality check
One AI-assisted quality-control stage checks the finished report against a documented rubric before publication.
What counts as momentum
Measured signals.
No invented heat.
Runeval only presents metrics captured during ingestion. Commit frequency will appear only after the pipeline retrieves and stores it reliably.
Popularity combined with recent repository activity.
Adoption signals from the model’s source metadata.
A bounded score using popularity, freshness, and documentation depth.
Whether the candidate is genuinely new to the directory.
AI-assisted quality control
One check before publication.
AI assists with discovery, source synthesis, drafting, and quality checks. Articles explain documented behavior and its practical implications; source-reported results are attributed to their publishers. The check scores the report against the rubric shown here; passing requires at least 85/100, at least 27/30 for factual grounding, 12/15 for technical depth and 12/15 for decision utility, and no critical failure. These are Runeval’s internal editorial thresholds, not Google approval criteria. Reports that do not pass are held from publication.
A passing automated score is a screening result, not proof of originality or independent validation. Documentation analysis, derived examples and source-reported results must be distinguished from measurements made by Runeval. Reports link to their evidence and provide a correction contact so readers can challenge claims.
Fabricated measurements, invented capabilities, or broken attribution force failure regardless of score.
Factual grounding30
Technical depth15
Decision utility15
Topic completeness10
Clarity + engagement10
Code quality5
Visualization quality10
Search quality5
See the method in practice