The problem

High reward can reflect a shortcut through a misspecified objective rather than intended behavior. In VGC doubles, voluntary switching is costly, discrete, and strategically revealing — if an agent has internalized transferable strategy, it should be visible in how, when, and why it switches. The question isn’t whether agents win; it’s what they actually learned.

Technical approach

A reproducible analysis layer around VGC-Bench and poke-env with project-owned switch-level logging at the policy-to-environment boundary. Two metrics: a 10-signal Strategic Switching Score over voluntary switches (matchup gain, damage mitigation, threat preservation, speed-control alignment, and six more), and an outcome-linked Residual Strategic Rate. Five agents evaluated across a locked training team and three transfer teams — 1,000 battles per cell, with Mann-Whitney comparisons under multiple-comparison correction.

Technical highlights

• The cautionary case, quantified: pure self-play (5.0M timesteps, PPO) holds a ~0.40 fixed-opponent win rate while switching almost never — win rate is not a proxy for learned competence.
• Behavior-cloned agents show the clearest interpretable strategy in-distribution; everything degrades under team transfer.
• Honest reporting: the preregistered exact-horizon metric is disclosed as underpowered; the residual analysis is labeled post-hoc.

Why it matters here

For ML/AI research audiences: RL evaluation methodology — behavioral probes alongside reward metrics, with the statistics to back the distinction.