Imagine-then-act world-model loops improve fast in imagination, but their realfrontier lift is bounded by independent contact with reality. This paperintroduces the evidence frontier, the best real competence supported by thecurrent independent real-contact ledger; Reality-Bounded Improvement (RBI),the claim that frontier lift is bounded by new independent real-contactinformation; grounding efficiency eta; Value of Real Contact (VoRC); and theGrounding-Efficiency Engine (GEE), which allocates scarce real contacts todecision-relevant residuals. The central correction is explicit: zero new realcontact does not mean zero learning, because systems can catch up to oldevidence. The sharp claim is that zero new independent contact cannot reliablyraise the evidence-supported frontier. A constructed witness over 150 seedsshows imagined score rising without frontier lift, and shows GEE outperformingrandom, coverage, risk-only, and relevance-only contact allocation at equalbudget. The package also includes frozen cloud return packages for V-JEPA2 andOpenVLA initial-observation proxy scoring plus a UnifoLM full-rollout payloadwith V-JEPA2 feature scoring. These cloud returns are bounded artifacts:the strongest current tier is OPEN_MODEL_RSHADOW_FULL_ROLLOUT_CANDIDATE,not open_model_rshadow_completed=true. No real robot, third-partyreplication, live deployment, production claim, or unconditional GEE optimalityclaim is made. It also does not claim priority over active learning, value ofinformation, experimental design, or submodular greedy approximation.