Each conditional probe submits a real, signed transaction carrying a precondition that cannot hold — a storage root that was never that value, a block window that closed a hundred blocks ago. If the sequencer enforces the guard it refuses the submission before inclusion, and a refused conditional costs nothing at all. So a probe only ever spends gas when the chain fails to enforce something it claims to, which is precisely the case worth paying for.

The read-only probes cost nothing either way. Nearly all of them are put to the chain operator’s public RPC rather than to ours, because the question is what somebody gets when they use the endpoint they were given — and because our node refuses to accept transactions at all by configuration, so aiming a conditional at it would measure our own settings and call the result the sequencer’s behaviour. The one row aimed at our node is there to hold our archive claim to the same test as everybody else’s.

An unreachable endpoint is recorded as no answer and counted as neither agreement nor contradiction. It also cannot write to the change history: a bad hour at an RPC provider must never be published as “the chain stopped enforcing this, then started again”. On a page whose only claim is that it reports what it measured, inventing an event would be the one unforgivable bug.

What accumulates is the interesting part. Any single answer here is dull and repeats every half hour; the record of when an answer changed is what turns “this chain does not enforce that” into “this chain stopped enforcing that on a date” — a statement somebody can act on.