What the last run measured.

How this is measured

• The suite runs real calls against the deployed endpoint, not a mock: registry lookups, VAT checks, IBAN validation, sanctions screening.
• A scenario passes only if the answer is usable by an agent: right shape, no leaked internals, and errors that say what to do next.
• Failure cases count. Unknown companies, malformed input and unsupported countries are part of the run, because that is where tools usually break.
• p95 is the slowest response in the top 5 percent, measured end to end from outside Cloudflare, so it includes the upstream registries and job boards. Searching many job boards at once means the slowest board sets the number, and it moves between runs.
• Some sources throttle or go down without warning: the European VAT service does it regularly. When that happens the scenario is marked degraded upstream and still counts as met, because what we promise in that case is to fail well: a clean, actionable, retryable error rather than a break or a wrong answer. It is counted separately and shown in its own figure above, so a run where an upstream was down never reads as a run where everything worked. If the tool had broken or leaked internals instead, it would count as our failure.