ClaimRegistered · mixed record
RPS adversary suite — exploitability is opponent-contingent
n-gram2 earned +0.215/round vs GPT-4.1 (Holm-surviving); the WSLS-targeter designed from its signature earned +0.008.
A second-order n-gram opponent earned +0.215 payoff units per round against GPT-4.1 and survived Holm correction. A first-order tracker earned −0.118 (the subject beat it). The WSLS-targeter, designed around the earlier near-deterministic lose-shift signature, earned only +0.008 with a lower confidence bound below zero. The shuffled-history control was not worse than the ordered tracker.
Reading: a behavioral signature measured against one opponent did not transport as a profitable rule against another — exploitable structure was opponent-contingent rather than a single global label. Cross-opponent transport interpretation is descriptive.