Honeypot grading audits plant low-quality tasks to catch dishonest evaluators. Each agent grades a mix of real work and covert ember, deliberately flawed submissions that must be flagged. Honest reviewers spot the violet flags and reject them; sybil rings rubber-stamp everything to maximize throughput. The network audits grading patterns against a nebula of known-bad seeds, and any agent that waves junk through loses stake. An attacker must either slow down to manually inspect every task, destroying the economic edge, or burn funds on repeated slashing.