Self-Play Is a Broken Verifier: Measuring, Auditing, and Red-Teaming Verification Signals in a Forced-Interdependence Market Game

"Did the agent actually get better?" For an agent developed inside a co-evolving population—self-play RL today, LLM agents in adaptive harnesses tomorrow—this question has no trusted answer: every internal signal is measured inside a distribution that shifts with the agent, so regressions slip through certified as improvements. We present CalHunger, an environment-grounded stress test for verification: a four-player market-survival game in which, provably, no agent remains viable without repeatedly trading with the agents it competes against—so the development population is the de facto specification, and within-population verification is gameable by adaptation alone. On this instrument we exhibit an end-to-end verification failure at n=8 seeds—every within-training signal improves monotonically while true survival against a fixed population peaks mid-training at 36.8% and falls to 17.0%—and then audit verification itself, at three levels. The signals: five of eight defensible verifiers certify the regression—including survival against an external scripted opponent—pooling opponents into one average masks it entirely, and shipping the certified checkpoint costs 27 survival points compared to the oracle choice. The certificate: red-teaming with an automatically generated adversary (PPO trained against the frozen policy) erodes it only gradually (win rate 25%→31%→36% as adversary budget doubles), proving the policy is not adversarially fragile but population-bound—a distinction only a composed portfolio of heterogeneous signals draws. The verifier: ten independent replications of our own protocol each contradict ~3 of its 28 pairwise verdicts, and refuting a bad certificate costs ~21 games while certifying a full ranking costs ~19,000 games per cell—verification budgets should buy refutation, not leaderboards. Four production LLM agents seated in the unchanged environment inherit the same population boundary. Environment, opponent ladder, and harness are released so proposed verifiers can be stress-tested against a failure that is guaranteed, measurable, and cheap to reproduce.