FAQ

Published rankings are generated from pipeline outputs and documented methodology, not vendor marketing claims.

For credible releases, keep the snapshots in src/data/db/site.sqlite synchronized with tasks, weights, dates, and reproducibility notes in Methodology.

Static HTML favors SEO, time-to-first-byte, and global CDN caching. Leaderboards may refresh on a daily or weekly cadence via CI-triggered builds.

Live queries may use read-only edge APIs alongside static snapshots and source citations for auditability.

The path outside the locale prefix is preserved (e.g., /zh/model/ ↔ /en/model/) for side-by-side reading.

Untranslated long-form sections may temporarily mirror English or another default—incremental localization.

Composites aggregate metrics after normalization and weighting—useful for overview, inadequate alone for weak-task analysis. Production deployments should publish per-task scores or sub-ranks where applicable.

When mixing public benchmarks, declare versions and missing-cell handling.

Model and agent boards emphasize capability or task success; GitHub trends emphasize OSS activity—they are complementary.

High star counts do not imply state-of-the-art capability; closed-source or off-GitHub work is excluded from trend statistics.

Typical flow: run evaluators in CI, write boards/datasets into src/data/db/site.sqlite via the data pipeline, trigger Astro build, deploy static output.

Object-storage snapshots require URLs and checksums on the Sources page.

External links open in a new tab with noopener/noreferrer. Trustworthiness and privacy policies of destination sites are the visitor’s responsibility.

Upstream data and code licenses apply; republish with Methodology and Sources links and the data date.