Methodology Changelog
Every change to our scoring formula is versioned and documented here. When we update weights, add new signals, or change normalization, we increment the version so you can track exactly how rankings are computed.
Current version: v2.1.0
Released February 25, 2026 - Added expert panel signal at 0% initial weight
Added expert panel signal at 0% initial weight
Introduced the Expert signal dimension to the scoring formula. It launches at 0% weight so existing scores are unaffected, but the infrastructure is ready to gradually increase the weight as expert reviews accumulate.
Changes
- New "Expert" signal added to all category weight tables with 0% initial weight.
- Expert signal half-life set to 90 days (longest of all signals).
- Infrastructure prepared for future structured expert reviews.
- Expert signal will be activated once the review program launches.
Category-specific weights introduced
Major methodology overhaul: each top-level category (Coding, Image Generation, Video Generation) now has its own weight profile. Previously, all categories shared the same weights. This change means the same model can receive different scores depending on which category it is ranked in.
Changes
- Coding weights: Quality 30%, Usage 28%, Demand 14%, Cost 14%, Trust 12%, Social 2%.
- Image Generation weights: Quality 26%, Usage 22%, Demand 18%, Trust 14%, Cost 12%, Social 6%.
- Video Generation weights: Quality 28%, Usage 21%, Demand 16%, Trust 15%, Cost 10%, Social 8%.
- Historical scores re-computed under the new weight profiles for continuity.
- Category-specific normalization pools: models are now compared only within their category.
Social signal capped at 8%
Added a hard cap on the Social signal to prevent viral hype or coordinated campaigns from dominating the composite score. Social sentiment remains a valid signal but is now bounded regardless of the raw value.
Changes
- Maximum Social signal contribution hard-capped at 8% of the composite score.
- Cross-validation layer added: isolated social spikes without corroborating signals are dampened.
- Social sentiment NLP pipeline upgraded with improved sarcasm and irony detection.
- Added monitoring alerts for abnormal social signal velocity (>5x baseline in 6h window).
Added additional demand signal pipeline
Integrated an additional independent traffic signal for the Demand category. This provides independent internet traffic data that complements existing search interest signals and improves demand signal accuracy.
Changes
- New traffic data pipeline connected and validated.
- Demand signal now blends multiple independent sources for more robust measurement.
- Added source health monitoring with automatic fallback if data is unavailable.
- Pipeline outputs validated against 30 days of historical data before going live.
Initial scoring formula launch
The first public release of the LM Market Cap scoring methodology. Establishes the composite score formula, the 6 original signals, and the normalization pipeline that converts raw data into 0-100 ModelCap Scores.
Changes
- Composite score formula: sigmoid(weighted_z_score) * 100, producing 0-100 range.
- Six signals launched: Quality, Demand, Usage, Cost, Trust, Social.
- Z-score normalization with clipping at [-2, +2] to bound outliers.
- EMA time decay with per-signal half-lives ranging from 7 to 60 days.
- Uniform weights (equal across all categories) as the baseline.
- Connected multiple independent data pipelines covering benchmarks, search interest, repository activity, package downloads, community sentiment, provider pricing, uptime monitoring, and academic leaderboards.