That table is category evidence but the real gap is operationalization, and the classifier still runs on expectedTerms.some() plus a byte floor, which means it will completely miss the akamai failure the prose already describes, a big empty shell with price and availability just gone. Cloudflare challenges and akamai denials need different next hops and separate retry budgets but the sample just treats both as continue. I would put cost per valid result, counting failed attempts, right next to pass rate and judge latency by p95 against the agent deadline instead of the mean. That scraperAPI number, 43s average with 19 of 24 requests over 30s, is already telling you this is a control flow problem and not a proxy quality one. 24 URLs is enough to separate cloudflare from akamai behavior, but 3 akamai pages alone can't size a real domain policy
That table is category evidence but the real gap is operationalization, and the classifier still runs on expectedTerms.some() plus a byte floor, which means it will completely miss the akamai failure the prose already describes, a big empty shell with price and availability just gone. Cloudflare challenges and akamai denials need different next hops and separate retry budgets but the sample just treats both as continue. I would put cost per valid result, counting failed attempts, right next to pass rate and judge latency by p95 against the agent deadline instead of the mean. That scraperAPI number, 43s average with 19 of 24 requests over 30s, is already telling you this is a control flow problem and not a proxy quality one. 24 URLs is enough to separate cloudflare from akamai behavior, but 3 akamai pages alone can't size a real domain policy