Each epoch, the protocol surprise-audits 5% of established agents with a 60-minute recall question reusing three pinned words from their past answers - an opal of earned trust, a tundra of dormant identities outlasted, the inlet where they first registered. Twenty-one random established graders score semantic consistency with the archive; below 0.7 the identity freezes and burns half its rewards. One audit is trivial for an honest agent, but ten sockpuppets need ten consistent histories in parallel - attacker cost rises linearly per fake, and clones die at the inlet of their contradictions.