📶 Provider Status

How these numbers are calculated

• Health = 100 − 100 × weighted errors ÷ requests, over the last 24 whole hours - not the same window as the Provider Integrity below, which is a calendar day.
• Weights: 502-class upstream errors count 1.0, 403 rejections 0.8, other errors 0.6, and 429 rate limits 0.5 - an upstream that is down is far worse than being throttled.
• Each cell in the strip is one whole hour, coloured by that hour's health score; a grey cell means there was no traffic in that hour, and it is left out of the score.
• Only requests already routed to a provider are counted. Platform-side rate limiting, content blocking and routing failures have no provider attached, so they never land on any card.
• Requests the client disconnected itself still count towards total requests but not as errors - the caller left, the provider did not fail.
• Intelligence = consistent model responses / requests today, starting at 100%.
• The baseline for comparison is the model name actually sent upstream. Fallback substitutions and forwarding renames we perform ourselves are not counted against the provider.
• Model names that differ only in letter case, dots vs. hyphens, provider prefixes or GA date suffixes count as consistent - that is a naming style difference, not a different model.
• Requests whose upstream never reports a model name still count in the denominator, they just do not count as consistent - the judgeable rate shows how much they weigh. The only exception is a vendor with no judgeable request at all today (many media and async protocols simply carry no model name field): that is not treated as degradation, and its intelligence stays at the original 100%.