A transparent record of the calls that shaped ModelRank. This is the project's "living research brain" between NotebookLM sessions. Every entry is reversible by the community via Discussions + PR.
- Decision: All growth work is genuine. No fake engagement, star farming, dummy issues, manufactured velocity, or deceptive marketing.
- Rationale: Deceptive tactics violate platform ToS and would destroy the one asset we have — independence/credibility. Real value compounds; fake velocity craters the moment it's noticed.
- Action: README number fixed (150 → 954 verified), one-line
curlbadge added, hero + feature grid + FAQ added to the UI, roadmap + this log committed.
- 5-dimension composite (Benchmarks 70 / Recency 15 / Community 10 /
Efficiency 5 / Reproducibility 0 reserved). Weights live in
config/settings.pyand are community-votable. - No paid placements for the score. Visibility can be bought if labeled; the composite score cannot.
- Independence is the product. We build no models, sell no weights.
- Raw data is public.
full_rankings_*.csv+leaderboard.jsonare the source of truth for every claim on the site and README.
- Should Efficiency rise from 5% → 15%? (Discussion pending)
- Which benchmarks to add next as the verified set grows?
- Is a cost dimension worth the data-quality risk?