Methodology & signal accuracy
QYNION combines five normalized factors into a regime-weighted alpha score. Each factor is calibrated to the current ecosystem regime for accurate signal detection.
Five factors
- Adoption Momentum — developer activity, model and package downloads, hiring and developer Q&A. Each signal is ranked against the other tracked entities first, so no single signal's scale dominates; weights are shared over the signals an entity actually has. A signal that doesn't apply (for example, open-model downloads for a lab that publishes no open models) is left out rather than scored zero.
- Developer Velocity — developer ecosystem activity and repository momentum signals
- Sentiment Trend — news sentiment direction across multiple intelligence feeds with fallback coverage
- Performance Delta — model capability from live sources: independent quality ratings (overall, coding, math) on a fixed scale, context window and vision support, and whether weights are open. A curated baseline fills gaps only, and each score lists the kinds of data it used. Entities that don't make models are not scored on this factor: its weight is shared over the other four.
- Attention Flow — attention signals combining developer activity, media coverage, and research output
Which provider feeds each signal, and full methodology documentation, are shared with Enterprise customers under NDA.
Published accuracy (30d)
Loading accuracy report…
Signal confidence
Once a signal type has 30 or more measured 7-day outcomes, its confidence is the lower bound of the 95% interval of its measured hit rate, so it reflects the track record rather than a fixed number. Until then a stated prior is used, and the signal says so.
Methodology changes
- October 2026: performance now uses live capability data for every model maker and no longer scores entities without models; adoption compares each source on its own scale; signal confidence is measured. Scores for frameworks and infrastructure rose (they were scored zero on model capability), and adoption spreads out more evenly across sources.
- September 2026: stale and implausible inputs are left out of scores, every score carries its coverage, and bearish signals are graded on a fall.
Narrative hot streaks
Macro theme persistence is counted in UTC calendar days, not pipeline run count. A theme must stay at BUILDING saturation or above to extend its streak.