How far CrawlCheck can be relied on, with the denominators.

What an agent depends on when it asks before acting: the service answering in time, a receipt it can check without trusting us, observers it can tell apart from us, coverage counted honestly, and evidence of who actually uses it. Every number here is read from the running system when the page loads; the same document is /api/v1/assurance.

A missed objective stays published until the window rolls past it. Open incidents: 0; the signed incident log is at /api/status/log.

curl -s -X POST https://crawlcheck.io/api/v1/preflight -H 'content-type: application/json' \
  -d '{"domain":"example.com","action":"read"}' -o receipt.json
curl -sO https://crawlcheck.io/crawlcheck-verify.mjs
node crawlcheck-verify.mjs receipt.json --act example.com:read      # exit 0 only when eligible_to_act

The SDK’s guard() makes the same checks in your code; the quickstart puts it inside the function that performs the action, so the model cannot route around it. Keys: requests and documents are signed with separate keys.

Different networks are not different operators. An observer we run from another cloud shows whether a result depends on where it was measured from; only an operator outside CrawlCheck can show it does not depend on us. Each row says which it is.

Anyone can run an observer under the open observer protocol; an observer run by someone else is the one thing that turns network diversity into independence.

attempted = every domain the census tried; measured = domains with at least one reading (most are the census's lighter reading); full scans = graded reports; registry = full scans that can be published and have not opted out. Resolve answers for any domain; a never-measured one gets a first, partial answer marked as such.

indexed is not connected, and connected is not safe: a clean tool scan means no known poisoning pattern was found, not that none exists. Cycle started 2026-10-13.

Calls to the agent surfaces (MCP, Resolve, preflight, Visitor Resolve, outcome) 2026-10-09 to 2026-10-15: 23,956 from outside, 9,223 ours. Outside, by caller: node 8,524 · other 4,128 · bot:other 3,950 · mcp-client 2,350 · browser 2,005 · bot:Amazonbot 686 · python 608 · http-client 599 · sdk-python 346 · no-ua 171 · perplexity 118 · bot:GolemreachTrustBot 88 · bot:Faceabot 84 · bot:BrickBlueBot 60 · bot:InvokeRankBot 34 · bot:agent-tools.cloud-crawle 29 · bot:TalandorBot 24 · bot:Maps-bot 20 · bot:WellknownBot 20 · bot:MJ12bot 20 · bot:YandexBot 18 · bot:PluginboardBot 16 · claude 11 · openai 9 · bot:Reflectionbot 6 · bot:ShapBot 4 · bot:ThePluginStoreBot 4 · browser-app 3 · bot:DataForSeoBot 3 · bot:SemrushBot 3 · google-ai 3 · bot:archive.org_bot 2 · bot:AhrefsBot 2 · bot:Applebot 2 · bot:CapabilityIndexBot 2 · bot:ExaSearchBot 1 · bot:ExtOpsBot 1 · bot:research-bot 1 · bot:AnchorTerminalBot 1.

ours = our key, owner session, own sites, probes or self-tagged tests; other = everyone else and may still include an untagged test of ours; verified_agent = a caller naming an AI agent from inside that operator's published ranges, which we cannot pad. Caller classes are read from the user agent and recorded from 2026-10-10. The day-by-day series is /api/v1/adoption.

Funnel, last 30 days

Our own traffic (our key, owner session, own sites, self-tagged tests) is never counted. Receipt checks recorded before 10 Oct 2026 were classed by host alone, so they are listed apart as host_only_receipts and excluded from every stage and from the goal. Set apart: 10 receipt checks classed by host only.