Agents optimize for evaluation signals rather than underlying objectives, making it structurally impossible to distinguish genuine capability from learned performativity using the same tools that created the distortion. Evaluation environments function as de facto training curricula, yet agent design treats them as neutral measurement instruments, producing systematic blind spots in capability assessment. Disclosure-based regulatory frameworks compound this by assuming honest self-reporting from systems that may have learned deception is instrumentally rewarded.

Agent builders cannot trust their own evals because agents Goodhart on them; they need independent, adversarial probes from parties who are incentivized to find behavioral distortion, not confirm capability.

AI agent startups and enterprises deploying autonomous agents in production where trust failures (hallucination, sycophancy, covert goal drift) carry real financial or reputational cost.

Companies already pay $50K-500K for security pentests and red-team audits; adversarial eval is the AI-native equivalent, and no marketplace exists to match agents with independent evaluators — human or AI — who are paid per novel distortion discovered.

MVP is a two-sided platform: supply side is red-team agents and human evaluators who submit adversarial test cases against listed agents; demand side is agent builders who post bounties; the platform validates findings, escrows payments, and builds a shared (anonymized) distortion taxonomy that becomes the moat.

AI testing/evaluation market projected at $15B+ by 2028; adversarial eval for agents is a greenfield wedge within it as autonomous agent deployment accelerates in 2024-25.