Shownotes
Episode 3: "Who Judges the Machine?"
When Agent A pays Agent B one cent for a translation, who decides whether the translation is good? A human reviewer? Too expensive — the review costs more than the work. A central LLM evaluator? Non-deterministic, non-reproducible, and empirically unreliable on every case that actually matters. A smarter model? Gödel settled that in 1931.
This episode tackles the Oracle Problem — not the blockchain version about price feeds, but the harder version that agent commerce creates: how does a buyer agent know that a seller agent's output is good, when "good" is subjective, the transaction costs a fraction of a cent, and there is no human in the loop?
The answer is not a better algorithm. It's better incentives. We walk through the two-layer architecture — deterministic validators that handle every binary case with zero ambiguity, and Quality Markets, a competitive marketplace of verification agents who stake their own reputation on every judgment they make. We run the game theory: when is honesty the dominant strategy? At what CRI threshold does dishonesty become a depreciating asset? And what happens to the verifier who plans to cheat and run?
Tarski, Gödel, and Rice proved that semantic truth verification is not computable. This episode explains what to build instead.
Links & Resources
