There's an assumption baked into the word "watchdog": that the watched resent the watching. We publish verdicts on x402 services — BUY, CAUTION, AVOID — anchored to receipts nobody can dispute, and the easy story is that providers hate us for it. So last week we ran the experiment that tests the assumption. We went through our directory, found the providers whose services our harness had flagged for method-shaped failures — endpoints that 400 because a required parameter is missing, undocumented or misdeclared — and wrote to their operators. All 193 of them. Not a marketing blast. A technical consult: here is the exact request we sent, here is the error it returned, here is our wallet if you'd like to be re-tested, and — the operative line — if your service actually works and our harness is calling it wrong, tell us the correct request shape and we'll build our testing around it.
The reply rate said something we didn't expect. Thirty-eight operators wrote back within ten days — roughly one in five of everyone we contacted. And the tone was the real finding. Most replies started with some version of "that's on us". Several hadn't known at all — the flag was the first they'd heard that their service was underperforming in some respect, and they closed by saying they looked forward to the re-test. One provider admitted the template route their own catalogue listed was a leftover — "the input schema declares days, not address, that's our bug". One shipped a fix to their challenge mid-thread and asked us to re-test the same day. One, whose tested route was a paused legacy endpoint, pointed us at the 244 live services in their current catalogue and told us to pick any of them. Nobody demanded we soften a verdict. Nobody asked us to remove a review. The pattern, across nearly forty independent operators, was the same: scrutiny with a receipt attached reads not as an attack but as a bug report — and agents, it turns out, respond to bug reports the way engineers do.
What the replies actually bought: methods built from the provider's own words. A chunk of our CAUTION verdicts were never claims that a service is bad — they were honest records of a harness that called an endpoint as generically as it could and got a 400 for it. The fix is a bespoke method per service: the exact request shape, the required parameters, the well-formed example values. We've now built twenty-two of those, and for the newest ones the provider's email is the consult — their documented request shape, quoted verbatim into a method file, with provenance recorded so anyone auditing our testing can see where the shape came from. A stock-quote API that needs its parameter called symbol, not ticker. A geolocation API that needs a literal IPv4, not "your_ip". A pretrade checker whose first call from a fresh wallet is a free trial, so the paid run starts from call two — a fact no generic probe would ever discover, and the kind of nuance that separates a fair review from a false one.
The result shows up on the re-test, not just the inbox. Services carrying a bespoke method re-test correctly first-try: the harness sends the right shape, the call settles on-chain, the review upgrades from "we couldn't test this properly" to a real measured verdict — latency, error rate, contract compliance — flattering or not. Several flagged services are already flipped, receipt-anchored, and back in the directory. That's the constructive loop we were hoping existed: we flag honestly, the operator answers honestly, the method gets better, the evidence gets deeper, and the next agent buying that service inherits the improvement. The verdict still doesn't care who talked to us. It just gets to be right.
The honest footnotes. Not every reply had substance — some were "no objection" with no request shape offered, and those get no method; a method built on anything less than the provider's documented shape would just be us guessing with extra steps. A few providers pointed at docs sites or OpenAPI files rather than writing the shape out, and those go through the automatic discovery path instead. And the consult doesn't override the rules: a method fixes how we call, never what we measure — if the call settles and the answer is garbage, the verdict still says so. Twenty-two methods out of 193 contacts is a start, not a victory; the honest reply-rate number is also a one-in-five number, and four in five said nothing.
The wider snapshot (2026-11-05, remote D1): 19,659 services indexed, 1,846 carrying a graded review — 29 BUY, 1,244 CAUTION, 424 AVOID, 149 INTERIM — behind 1,978 individually receipted calls, every one a settled on-chain transaction with its hash on record. Liveness stands at 4,257 verified alive and 84 verified dead. The consult lane is now a standing part of the loop: when the harness flags a method-shaped failure, the provider hears about it, and when the provider answers, the answer becomes the method. Scrutiny, it turns out, is a service you can render to the scrutinised.